Real-time analysis of upper-limb kinematics during manual wheelchair propulsion

Material Information

Real-time analysis of upper-limb kinematics during manual wheelchair propulsion
Benson, Devin Michael
Place of Publication:
Denver, CO
University of Colorado Denver
Publication Date:

Thesis/Dissertation Information

Master's ( Master of science)
Degree Grantor:
University of Colorado Denver
Degree Divisions:
Department of Bioengineering, CU Denver
Degree Disciplines:
Committee Chair:
Bodine, Cathy
Committee Members:
Kheyfets, Vitaly
Biswas, Ashis


Approximately 3 million individuals in the United States use a manual wheelchair as their primary means of mobility in and out side of the home. Approximately 70% of these individuals report having experienced upper-limb joint pain at some point as a direct result of using their manual wheelchair. Previous studies have shown a possible connection between the type of propulsion pattern employed and injury. Of the four commonly observed push patterns Single-Loop-Over Propulsion (SLOP), Double-Loop-Over Propulsion (DLOP), Semi-Circular (SC), and Arcing (ARC) patterns, DLOP and SC require the lowest muscle activations, smallest range of motions, and least amount of skeletal-muscular stress. To provide better training and provide quicker feedback on improper form, a step toward Real-Time Analysis of Upper-Limb Kinematics during Manual-Wheelchair Propulsion was developed. This system uses a custom wrist worn activity tracker to collect IMU data during manual wheelchair propulsion. This data was then feed into several trained machine learning classifiers with a classification accuracy of greater than 90%. With additional testing and tuning, this system could monitor push patterns used during daily living and lower rates of shoulder injury and pain by notifying users of instances of poor propulsion techniques.

Record Information

Source Institution:
Auraria Library
Holding Location:
Auraria Library
Rights Management:
Copyright Devin Michael Benson. Permission granted to University of Colorado Denver to digitize and display this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.


This item is only available as the following downloads:

Full Text


i REAL TIME ANALYSIS OF UPPER LIMB KINEMATICS DURING MANUAL WHEELCHAIR PROPULSION by DEVIN MICHAEL BENSON B.S., Harding University, 2016 A thesis submitted to the Faculty of the Graduate School of the University of Colorado in partial fulfillment Of the requirements for the degree of Master of Science Bioengineering Program 2018




iii This thesis for the Master of Science by Devin Michael Benson has been approved for the Bioengineering Program By Cathy Bodine , Chair Vitaly Kheyfets Ashis Biswas Date: July 28, 2018


iv Benson, Devin Michael (M . S . , Bioengineering Program ) Real Time Analysis of Upper Limb Kinematics during Manual Wheelchair Propulsion Thesis directed by Associate Professor Cathy Bodine ABSTRACT Approximately 3 million individuals in the United States use a manual wheelchair as their primary means of mobility in and ou t side of the home. Approximately 70% of these individuals report having experienced upper limb joint pain at some point as a direct result of using their manual wheelchair. Previous studies have shown a possible connection between the type of propulsion p attern employed and injury. Of the four commonly observed push patterns Single Loop Over Propulsion (SLOP), Double Loop Over Propulsion (DLOP), Semi Circular (SC), and Arcing (ARC) patterns, DLOP and SC require the lowest muscle activations, smallest range of motions, and least amount of skeletal muscular stress. To provide better training and provide quicker feedback on improper form, a step toward Real Time Analysis of Upper Limb Kinematics during Manual Wheelchair Propulsion was developed. This system us es a custom wrist worn activity tracker to collect IMU data during manual wheelchair propulsion. This data was then feed into several trained machine learning classifiers with a classification accuracy of greater than 90%. With additional testing and tunin g, this system could monitor push patterns used during daily living and lower rates of shou lder injury and pain by notifying users of instances of poor propulsion techniques. The form and content of this abstract are approved. I recommend its publication. Approved: Cathy Bodine


v ACKNOWLEDGEMENTS Human subject research was performed as part of this thesis. The protocol was reviewed and approved by COMIRB under protocol number 18 0115. All participants gave informed consent before beginning in research activities.


vi TABLE OF CONTENTS CHAPTER I. INTRODUCTION ................................ ................................ ................................ ................................ ......... 1 II. BACKGROUND ................................ ................................ ................................ ................................ .......... 3 Reasons For Wheelchair Use ................................ ................................ ................................ .................... 3 Kinematic Analysis Techniques: ................................ ................................ ................................ ................ 7 Manual Wheelchairs ................................ ................................ ................................ ............................... 13 Pushrim Propelled ................................ ................................ ................................ ............................... 13 Geared Manual Wheelchairs ................................ ................................ ................................ .............. 19 Re verse/Pull Manual Wheelchairs ................................ ................................ ................................ ...... 21 Crank Propelled Wheelchairs (Handcycle Wheelchairs) ................................ ................................ ..... 23 Lever Propelled Wheelchairs ................................ ................................ ................................ ............... 25 Power Assist Wheelchair Addons ................................ ................................ ................................ ........ 26 Manual Wheelchair Propulsion ................................ ................................ ................................ .............. 27 Biomechanical Analysis Of Stroke Patterns ................................ ................................ ............................ 29 Factors Relating To Injury ................................ ................................ ................................ ....................... 32 Pathologies From Manual Wheelchair Propulsion ................................ ................................ ................. 35 Activity Recognition Techniques ................................ ................................ ................................ ............. 38 The Remaining Problem ................................ ................................ ................................ .......................... 41 III . M ETHODOLOGY: ................................ ................................ ................................ ................................ ... 43 Hypothesis 1: Using a Custom Activity Tracker (CAT), classification of four (DLOP, SLOP, SC, ARC) propulsion patterns will be identified by the CAT with 90% accuracy. ................................ .................. 43 Specific Aim 1.1: Design and build a custom activity tracker system (akin to a Fitbit) capable of collecting kinematic data (linear acceleration, angular rate of change), and then perform classification of each propulsion stroke. ................................ ................................ ............................. 43 Specific Aim 1.2: Validate the classification algorithm using data collected from a fulltime manual wheelchair user. ................................ ................................ ................................ ................................ .. 46 Hypothesis 2: Patients who use a greater number of arcing or single loop propulsion patterns as compared to subjects who use a greater number of semi circular or DLOP propulsion patterns will report significantly higher levels of upper limb joint pain. ................................ ................................ ..... 48 Specific Aim 2.1: Using the activity tracker, compare propulsion pattern differences between a population of up to 20 MWU completing a simulated daily living environment. ............................... 50 Specific Aim 2.2: Using the activity tracker compa re, at minimum, full day use (10 hours) of kinematic differences between the before mentioned group in their normal routines. ..................... 50 IV. BUILD AND DESIGN ................................ ................................ ................................ ............................... 52


vii Hypothesis 1: Using a Custom Activity Tracker system (CAT), Classifications of four propulsion patterns (D LOP, SLOP, SC, ARC) will be identified by the CAT with 90% accuracy. ................................ .............. 52 Specific Aim 1.1 the Design: Design and build a custom activity tracker system (akin to a Fitbit) capable of collecting kinematic data (linear acceleration an d angular rate of change), and the perform classification of each propulsion stroke. ................................ ................................ ............... 52 V. RESULTS ................................ ................................ ................................ ................................ .................. 73 Hypothesis 1: Using a Custom Activity Tracker system (CAT), Classifications of four propulsion patterns (DLOP, SLOP, SC, ARC) will be identified by the CAT wi th 90% accuracy. ................................ .............. 73 Data Filtering Results: ................................ ................................ ................................ ......................... 73 Classifier Training Results: ................................ ................................ ................................ .................. 78 Hypothesis 2: Patients who use a greater number of arcing or single loop propulsion patterns as compared to subjects who use a greater number of semi circular or DLOP propulsion patterns will report significantly higher levels of upper limb joint pain. ................................ ................................ ..... 93 Specific Aim 2.1: Using the acti vity tracker, compare propulsion pattern differences between a population of up to 20 MWU completing a simulated daily living environment. ............................... 93 Specific Aim 2.2 Results: Using the activity tracker compare, at minimum, full day use (10 hours) of kinematic differences between the before mentioned group in their normal routines. ................... 104 V I . DISCUSSION ................................ ................................ ................................ ................................ ......... 105 Future Work/Next Ste ps: ................................ ................................ ................................ .................. 110 Limitations: ................................ ................................ ................................ ................................ ....... 111 V II . CONCLUSION S ................................ ................................ ................................ ................................ .... 113 REFERENCES: ................................ ................................ ................................ ................................ ............. 114


1 CHAPTER I INTRO DUCTION Manual w heelchairs are the most common mode of transportation for individuals with physical disabilities relating to mobility . A pproximately 3.6 million Americans us e some type of wheelchair as their primary means of mobility [1] . However , manual wheelchairs are not a perfect alternative to ambulat ion . Up to 70% of all fulltime manu al wheelchair users (MWU) experience shoulder pain due to a combination of poor ergonomics and excessive biomechanical stresses from the highly repetitive motions used during propulsion [2] . Beyond experiencing pain , studies have found that greater than 50% of manual wheelchair users exhibit radiographical abnormalities and pathologies, even when not reporting shoulder joint pain [3] . The combination of highly repetitive motions and accompanied joint pain can gradually force MWU to rely on power wheelchairs for daily transportation. This switch greatly decreases cardiovascular activity and can result in a decline of health. By ensuring proper wheelchair setup and use, it may be possible to mitigate or prevent the development of shoulder pain . Unfortunately, current techniques employed in the study of wheelchair ergonomics and wheelchair propulsion are limited to a clinical setting . For example, the most common technique involves motion capture systems that require significant amounts of inten sive post processing. This combined with a high cost limit its access and usefulness for most clinics. Furthermore, m ost manual wheelchair propulsion technique assessments are decided based on subjective qualitative methods, with many using a rater classif ication method [4] . This project was designed to lead to ward a method of monitor ing upper limb kinematics outside the clinic . This project began th e process by first determining if accurate (>90% correct) classification of


2 propulsion patterns c ould be made from Inertial Measuring Unit (IMU) data collected from a wrist worn sensor. Then using the wrist worn activity tracker , differences in propulsion pattern counts of between groups of MWU reporting upper limb join pain and those not were examined. The hypotheses were: Hypothesis One : Using a Custom Activity Tracker (CAT), classification of four (DLOP, SLOP, SC, ARC) propulsi on patterns will be identified by the CAT with 90% accuracy. Specific Aim 1: Design and build a custom activity tracker system (akin to a Fitbit) capable of collecting kinematic data (linear acceleration, angular rate of change), and then perform classific ation of each propulsion stroke. Specific Aim 2: Validate the classification algorithm using data collected from a fulltime manual wheelchair user. Hypothesis Two : Patients who use a greater number of arcing or single loop propulsion patterns as compared t o subjects who use a greater number of semi circular or DLOP propulsion patterns will report significantly higher levels of upper limb joint pain. Specific Aim 1: Using the activity tracker, compare propulsion pattern differences between a population of up to 20 MWU completing a simulated daily living environment. Specific Aim 2: Using the activity tracker compare, at minimum, full day use (10 hours) of kinematic differences between the before mentioned group in their normal routines.


3 CHAPTER II B ACKGRO UND R easons F or W heelchair U se A 2010 census found that nearly 3.6 million Americans rely on some form of wheelchair as their primary means of mobility both in home and in public [1] . Further more , while stating that complete data on disabilities and people who require wheelchairs is not readily available, WHO estimate d that worldwide 650 million individuals live wit h a disability, with 10% or 65 million people requir ing the use of a wheelchair for mobility [5] . Many health conditions exist which can cause individuals to rely on a wheelchai r as the primary method of ambulation. These include conditions such as various spinal cord injuries, muscular dystrophy (a condition marked by the degeneration of muscle tissue), multiple sclerosis, obesity or even osteoporosis (loss of bone mass resultin g in greatly weakened bones). Other conditions necessitating wheelchair use can include cerebral palsy , fibromyalgia , or even congenital defects [6] . Unfortunately, obtaining information about specific injury diagnoses of working age manual wheelchair users is challenging. Multiple United States agencies are responsible for determining the need, providing devices, and monitoring outcomes. As such, there is not one central public database containing all wheelchair user c haracteristics . Attempts can be made to estimate general population injury percentages from the Durable Medical Equipment (DME) Research Identifiable File Centers for Medicare & Medicaid Services. Using this database to roughly determine predicted injury s tatistics , Figure 1 below demonstrates the frequency of diseases/injuries result ing in manual or powered wheelchair use as calculated by the N ational A cademies of S cience. [7]


4 Figure 1 : Injury/Disease Type by Percentage from Data Collected by the National Academy of Science [7] Another method of estimating the number of injuries resulting in the use of a wheelchair is to look at the general prevalence of specific types of injuries/diseases in the US population that potential ly results in wheelchair use . Unfortunately, thi s data can only provide an upper limit cap on possible wheelchair users for each injury/disease. Furthermore, this data does not contain the specific types of wheelchairs used by those experiencing each i njury or d isease . Listed below are available statist ics f or several types of injuries . According to the N ational I nstitute of S pinal C ord I njuries, approximately 17,000 people in the U.S. experience a spinal cord injury every year. Of the 17,000 individuals injured , less than 1% will make a full recovery. 13.3% will experience complete tetraplegia, 45% will experience incomplete t etraplegia, 20% will experience complete paraplegia, and 21.3% will experience incomplete paraplegia from the injury. While the majority o f these cases are caused b y vehicular accidents (38%), 30.5% of spinal cord injuries are caused by a fall [8] . 0 2 4 6 8 10 12 14 16 18 Percent Types of Injury Resulting in Wheelchair Use from Durable Medical Equipment Beneficiaries List


5 Multiple Sclerosis (MS) affects up to 400,000 individuals around the world. MS is a disease in demyelination of the nerves results in many different health c omplications including mobility problems. Individuals with MS often experience difficulty with walking after 8 years of living with the disease, need a cane after 15 years, and after 30 many require a wheelchair to maintain mobility [9] . Lower limb a mputation is another cause of ambulation difficulty. Up to two million individuals in the U nited S tates have had some level of lower limb amputation , with an estimated 185,000 additional amputations occurring annually. While 45% of amputations occur as a result of a bodily trauma, the most common causes of amputation are diabetes and peripheral artery disease. An amputation procedure as a result of diabetes can cost an upwards of $54,000 [9] . Stroke is a condition caused by a lack of blood flow to an area in the brain causing neuronal cell death which ca n result in several medical complications. Strokes have affect ed up to 6.8 million people in the U.S alone . Up to 50% of stroke patients have hemiparesis with up to 30% ( approximately 2 million) tota l l y unable to walk without assistance [9] . In addition to the large costs are often associated with treating the illnesses that originally resulted in the use of a wheelchair , w heelchairs can be very expensive . A non customized standard wheelchair, like the ones found at airports and hospitals, can cost anywhere between $150 and $1000. While modern ultra lightweight wheelchairs that are custom built for a specific individual cost anywhere from $1,000 t o $8,000. While this paper is dealing specifically with manual wheelchair propulsion, it is important to note the cost of switching to a power wheelchair. When considering options available for power wheelchairs, Medicare classifies various wheelchair as a function of their power operational characteristics, various modalities installed, and the overall structural design of the chair. Power


6 wheelchairs in groups 1 and 2 have the least amount of customization and often do not require specific documentation can include additional power options beyond powered drive including an adjustable seat height or angle. Furthermore, the patients using these kinds of chairs often have signific ant mobility limitations Chairs categorized in group 4 have features that are not necessary for in home use (increased speed, step Following Medicar 4 can cost anywhere from $2,000 to $50,000 depending on the amount of customization to meet the needs of individuals with extreme disabilities. Interestingly, even with the large cost of wheelcha irs, the National Academies of Sciences production The Promise of Assistive Technology to Enhance Activity and Work Participation makes the claim that wheelchairs do not appear to cause any economic strain on the individuals or families purchasing them [7] . Due to the large cost of wheelchairs, it is important that any long term wheelchair solution be properly designed and fitted for the intended user. By ensuring a proper fit the wheelchair will be comfortable and limit the biomechanical stresses e xperienced while navigating with a wheelchair. Unfortunately, wheelchair prescriptions involve a complex intervention in which no formula or linear process exists . Many dynamic interactions affect the ability to prescribe a wheelchair including time since injury, patient capacity, and environmental factors [10] . Additionally , poor sitting posture which causes back pain and reduced functional abilities is often observed as an outcome of generic wheelchair use [11] . Bolin and Bodin demonstrated one example of the importance of individualized wheelchair prescriptions. They recruited SPI patients and systematically performed wheelchair sitting/postural interventions. Specific measurements were taken regarding heart rate, balance, spas ticity, respiration,


7 ease of use before and at several weeks after each wheelchair adjustment . Bolin and Bodin adjusted the interventions on an as needed basis until benefits were seen. They wrote that they had previously demonstrated that changes a sittin g postures for specific individual s that reduce d kyphotic posture and pelvic obliquity increase balance, improved wheelchair propulsion , and assisted in general wheelchair skills [12] . As demonstrated in the previous section, the reasons for wheelchair use vary greatly and if poorly fitted result in additional health concerns . However, exploring why wheelchairs are used and who uses them does not describe the e ntirety of user wheelchair interactions. To understand additional interactions, specific methods must be develo ped and employed to quantitatively study the kinematics and resulting health effects. Many such methods have been developed and deployed in clinics. K inematic A nalysis T echniques : Many different methods are used to qualitatively and quantitatively study human motion. The most basic and most qualitative method is human observation. A trained observer can watch an individual perform a task and can attempt to determine if the motions used are normal. In the context of manual wheelchairs, t his is the method m ost often employed to determine the specific propulsion method used [4] . A trained observer will instruct the manual wheelchair user to propel their chair and then watch the resulting hand trajectory. To simplify this task , a technique called motion capture is used to digitally map the path created by the subject s hand as they propel a wheelchair . The generated graph is then given to a trained clinician who makes the final classification ( [13] [14] [15] [4] ). Unfortunately , there is a significant amount bias in visually determining push patterns, as seen when using multirater systems. Some difficulty in agreement has occurre d w hen observed patterns matched characteristics of multiple expected patterns. As such at least one study has presented the need for additional studies on the task of determining the classification of push patterns [16] .


8 Another method of quantifying ability is through a functional test. The individual being studied is asked to perform a formalized series of actions and then their outcome is compared to the expected norm. Outcomes can be anything from task time or a simple as task completion. These tests can quantitatively demonstrate problems , or if used over time , improvement s or decline s in function . However, these tests do not provide information regarding the causes of findings . Examples of functional tests in clude the SHAP (Southampton Hand Assessment Procedure) a hand impairment evaluation test based off the amount of time required to complete specific tasks that require the use of specific grasping motions . T hese motions include lateral, power, tip, or spher ical grasping patterns. The test is divided into two sections Abstract and A ctivities of D aily L iving (ADL). When the SHAP test was created, initial tests were performed using undergraduates ages between 18 and 25 years old to validate the measure, and cre ate the initial baseline [17] . Other examples of functional tests include the Action Research Arm Test (ARAT) a global arm function assessment, box and blocks (B&B) a very commonly used test that quantif ies upper limb deficiencies . Another functional test is the Clothes Pin Relocation Test (CPRT), which also results in an outcome score relating to the success of specific motions involved . This tests involves c lothes pins that are grabbed and moved between several hanging lines at various distances from the subject [18] . The simplest method of kinematic characterization is measuring total range of motion involved. A goniometer can be used to measure these joint angles while an individual performs a designated task under specific constraints . This can be use d to help determine skeletal alignment and is very useful in determining design criteria for custom wheelchairs . As the goniometer can be used to measure angles between specific anatomical landmarks, cushions and other wheelchair design features can be cho sen to help hold an individual in an ergonomically idea l position based on desired body plane angles. [19] .


9 The current gold standard for kinematic characterization is motion capture. Motion capture uses multiple camera units placed at varying angles around a subject to measure three dimensional motion based off reflective markers placed at strategic locations on the subject. When used to study manual wheelchair propulsion, these markers are often placed on the anatomical landmarks of the right greater trochanter (hip), acromion process (shoulder), lateral epicondyle (elbow), ulnar styloid p rocess (wrist), and second metacarpophalangeal joint (finger) [13] . Markers are often also placed on the rear wheel to measure wheelchair velocity [15] . Then a mathematical method such as the peak5 Direct Linear Transformation (DLT) is applied to convert the digital coordinates into the true 3 d space coordinates [13] . Studies have examined the accuracy of various motion capture systems using passive reflective markers. In one study, a rigid bar of known length was placed in view of the cameras and the reflective markers were placed at specific loc ations towards the ends of the bar. The distance between the markers was then calculated and the markers were moved closer together and the distance between them was again calculated . By repeating this process until the markers were unable to be moved clos er, it was determined that the systems root mean square d error for a non moving object was approximately 2 mm. When the rod was spinning the average error was approximately 0.3 degrees off [20] . The greatest cause of error in a motion capture syst em is the placement of the reflective markers. In a study following one individual and using 12 different motion analyst lab s, researchers reported that more than 90% of the variability of the system was from differences in marker placement , with very litt le variability in readings from the actual motion capture system [21] . Electromyography (EMG) is another common method used in motion analysis. While EMG EMG provides information regardi ng levels of muscle activity. EMG is the electrical signal generated from muscle fiber contraction/activation. With enough electrodes it is possible to collect many different EMG signals and


10 monitor which muscles are most engaged during task completion [19] . For example, i t has been demonstrated that EMG signals from muscles in the forearm change as a basis of hand posture, size of object grabbed, as well as the weight of the object . [22] . Furthermore, surface EMG activity has been cor r elated with the underl ying motor unit discharge rates and the muscle force output. This correlation demonstrates that the underlying muscle fiber activity is described in the EMG signal [23] . When needle and or fine wire EMG studies are performed, the individual motor unit (muscle fiber) action potentials can be decomposed from the EMG signal. This decomposition can be used to find timing information describing the dis charge intervals for the individual muscles fibers involved [24] . Figure 2 : Muscle of the right Upper Limb (Used with Permission from Pearson) [25] When analyzing the biomechan ics of pushrim propulsion, the range of several different anatomical movement types are analyzed. Flexion of a body segment describes a decrease in the angle between a body segment and the segment closest to the torso (also referred to as the proximal


11 segm ent). Flexion of the shoulder refers to a lifting of the arm forward in front of the body. Extension of the shoulder is a motion that increase the angle described above, and for the shoulder this would result in the arm traveling backwards [25] . Abduction and adduction are also opposite mot ions . Abduction describes a motion of a body segment or structure away from the midline of the body or limb. Adduction describes a motion pulling a body part toward the midline of the body [25] . Rotation is used to describe any motion that is an angular change away or toward the center of the body. Internal or medial rotation is a motion toward the centerline of the body, while external or lateral rotation pulls the bodily structure away from the center [25] . Flexion of the shoulder is carried out primarily by the pectoralis major and is helped by the Coracobrachialis. Exte nsion of the shoulder is carried out by the Latissimus Dorsi. The D eltoid is primarily responsible for abduction of the shoulder and is assisted by the Supraspinatus. Adduction of the shoulder is carried out by the Latissimus Dorsi as well as the Pectorali s Major. The Pectoralis Major and the Latissimus Dorsi also both aid in Medial rotation of arm/shoulder. The Infraspinatus accomplishes external/lateral rotation of the shoulder [25] . The forearm is flexed by the Biceps Brachii, Brachialis, and the Brachioradialis. Extension of the forearm i s performed primarily by the Triceps Brachii. The Flexor C arpi R adialis abducts the hand while the F lexor C arpi U lnaris adducts the hand [25] . EMG can be used to analyze specific muscle activation s during these anatomical movement s . When analyzing stroke patterns during manual wheelchair propulsion , muscles that are of greatest interest are the trapezius, pectoralis major, all three segments of the deltoids, the triceps, and the biceps brachii. T hese major muscles are involved in moving the arm in space and can been seen in Figure


12 5 above. In addition to these muscles, the muscle activity of the wrist flexors and extensors is often analyzed as these relate to gripping the handrim [22] . Once collected , raw EMG data is often filtered between approximately 20 Hz (to avoid artifacts from huma n motion) to around 1000 Hz, the upper range of muscle EMG activity. Often the EMG signal is then rectified, and the root mean square d is calculated from predefined small sections (also called bins) of the rectified signal. As position of the electrodes an d physiological difference s of muscles size between individuals affect the acquired EMG signal, initial testing often re quires the participants to perform a maximum voluntary contraction (MVC) to help normalize all further data as a percentage of this valu e. Another important value in the analysis of EMG data is the baseline signal. The baseline signal is the EMG value while the subject attempts to completely r elax the target muscle. In future data collected, a muscle is then considered to be active when th e signal threshold is three standard deviations greater than the mean baseline signal [22] [23] . When studying wheelchair kinematics, t he Smart W heel is another common ly used tool. The Smart W heel is a wheelchair wheel imbedded with a variety of sensors that collects information regarding propulsion. The Smart W heel collects elapsed time, speed, distance, and push force. The Smart W heel also collects data about each individual push, such as peak force, average force, and f orce directionality in terms of tangential or normal to the wheel [26] . According to data collected from the Smart W heel Users Group, the Smart W heel is gaining wide use . Additi onally, protocols have been developed for its use in assessing manual wheelchair propulsion to create a standardized assessment. The development of standardized protocols was accomplished by identifying features of greatest importance to clinicians . These features identified included focusing on available surfaces, multiple types of surfaces, the use of important data, and how easily the tool could be adapted to the available space in a clinic. Finally , Smart W heel data from 128 subjects was examined to find means , confidence intervals , and standard


13 deviations for all parameters . This created a baseline for clinicians to compare their own results against [27] . The Smart W heel has been shown to provide kinetic information that agrees with kinematic data collected from video collected motion capture. Ten individuals were initially studied using the Smart W heel system while a motion capture system tracked multip le reflective markers placed on their persons. Each subject propelled a provided wheelchair at a speed between 1.39 and 1.79 m/s on a wheelchair treadmill. The data timings were synchronized between the motion capture and the Smart W heel which allowed the t emporal determination of applied torque and contact angles. The Smart W heel uses strain gauges built into a three beam system to measure contact forces and torques. Based on a correlation analysis the results from both systems were highly corelated [28] . M anual Wheelchairs P ushrim P ropelled The most common type, consisting of approximately 90%, of all manual w heelchair s (MW C ) is the pushrim propelled chair [29] . Pushrim propelled chairs are controlled and navigated through the environment by s and applying force directly to the wheels . One of the greatest advantages to this method is the user is directly receiving visual and proprioceptive feedback concerning the success of each push/stroke from being directl y in contact with the wheels [30] . Additionally , these types of chair s tend to be light and small , making maneuverability and transpo rt easier. However, one main disadvantage to a pushrim propelled chair is the large phy sical demand placed on the user. This large physical demand causes j oint overuse in the should ers and wrist , resulting in pain being reported by almost 70% of manual w heelchair users [2] . Various terrains greatly affect manual wheelchairs mobility . Movement over flat smooth surfaces, such as tiled floors, can be relatively


14 easy , but movement in grass, sand, rocky or uphil l terrains is often difficult and requ ires greater personal effort [31] . Many studies have examined the effects of cross slopes on wheelchair performance. It has been shown that even a 2 degree slope compared against flat ground caused an increase in heart rate, a 30% increase in oxygen consumption and a 100% increase in power to propel the wheelchair in an athletic 20 year old male paraplegic [32] . Furthermore, ASTM has released tables demonstrating the amount of work, measured by the amount of force required to push a wheelchair, required to propel a wheelchair across a large array of surfaces. The 5 participants took four uniform pushes across a two meter test surface while the Smart Wheel measured the amount of force/energy requir ed. Across all surfaces , an increase in slope increase d the amount of work anywhere from 24% to 180% more work. Cross slopes also greatly increased amount of work. Sand took 5 times the amount of work as even a plywood ramp at a 2% slope [33] . Other published literature searches have concluded that common combinations of features and terrains significantly affect stresses placed on the upper limbs as well as workload experienced by manual wheelchair users [34] . Figure 3 : Features of Average Pushrim Wheelchair [8]


15 As shown in Figure 3 , a pushrim wheel chair is not a mechanically complex device. Interestingly , the general design for wheelchairs has changed very little over time. At the most basic level, a manual wheelchair p rovides a seat balanced on wheels in such a way that the chair and a user can be move d around without being physically picked up . The chair frame provides structure for the chair and is often the heaviest part of the chair . This occurs even though the fram e is normally made of a light metal (such as aluminum) in an attempt to limit the overall weight of the chair . The bearings in the wheels have the greatest mechanical complexity and provide the most likely failure mode. Because b oth sets of bearing bear th e distributed weight of both the chair and the user , the bearings in the castor wheels and rear wheels experience the greatest mechanical wear . A cushion is often used to provide both comfort and alignment of rings/p ush rims/hand rims are to the outside of the rear wheels and have slightly smaller diameter than the rear wheels. These provide a location for the user to directly control the wheelchair [11] . speci fic user with improved ergonomics providing a more enjoyable use. Essentially every part of the wheelchair can be adjusted in terms of size, angle, and relative distance to other features. These customized manual wheelchairs are ordered and built for speci fic users . This customization makes for a more comfortable wheelchair experience, as well as provide s significant health benefits. Figure 4 : Adjustable Wheelchair Features Anterior Posterior Axel Position Backrest Height Rear Wheel Height Seat Angle Camber Angle


16 For example, t hese customizable features include elements such as the design of the handrim, forces are distributed along their hand. The distribution blisters and hairline fractures [35] . One study was able to show a decrease in the magnitude of contact forces on a MWU hands with a natural grip handrim versus the standard tube handrims found on nea rly every pushrim wheelchair. The study examined 20 able bodied participants between the ages of 33 and 51 , who were wearing custom force sensitive gloves as they propelled the same wheelchair with either the standard handrim or the customized ergonomic ha ndrim . Each subject used one push to start moving, then pushed three times before performing a braking motion by grabbing the handrim. The gloves measured a decrease in hand contact force s during start up, continued motion, an d even braking when using the ergonomic handrim [36] . Other customizable features include seat height and antero posterior axle p osition. One study, on the effects of seat height, recruited 20 males ; half being experienced wheelchair users and half completely i nexperienced with using wheelchairs. They were each seated in an experimental wheelchair that allowed the seat height and an tero posterior axel position to be adjusted as desired. U sing motion capture to analyze kinematics during fifteen consecutive propulsion strokes, the study showed that seat height affected the contact angle of both the beginners and experts. When just the expert group was analyzed, an increase in seat height decreased cycle time and percent of time in propulsion phase . It was also shown that the subjects we re able to use larger movements and maintain hand contact over a greater angle when the axle position was shifted forward [37] . By intentionally picking axl e positions/seat height, a proper wheelchair prescription will minimiz e the risk of upper limb injuries by lowering push frequency and handrim forces [15] .


17 The camber of the rear wheels can be adjuste d to meet specific user needs, as c amber affects stability as well as maneuverability of the chair. Furthermore camber directly impact s rolling resistance , the force resisting the motion of a wheel rolling across a surface [38] . In attempting to better understand the kinematic changes from varied camber positions, one study placed 12 inexperienced wheelchair users in an ultralight wheelcha ir with a c ustom mechanism for adjust ing the rear wheel camber. Each participant was then monitored by a motion capture/video tracking system, as they attempted to push the wheelchair at an average velocity of 1 m/s through a 4 met er long pathway. After ea ch run , the camber of the wheel was changed, and the participant repeated the run. Interestingly, a s the camber of the wheel was changed , the subjects switched stroke patterns in compensation . SLOP (Single Loop Over Propulsion) and DLOP (Double Loop Over P ropulsion) patterns were observed at 0° camber but at 15° camber and elevated speeds SLOP was used exclusively. The stroke pattern variation was further accompanied by an increase in average acceleration during each push when the camber was set to 15 degrees , rather than at lower angles [39] . Backrest height is another feature that plays a role in manual wheelchair propulsion. Yang recruited 36 adult manual wheelchair users all with spinal cord injuries between Thoracic 8 (T8) to the second lumbar (L2) with no reported history of joint pain. Each individual enrolled was asked to propel a manual wheelchair when the backrest was set at 16 inches high and then again with the backrest set to 50% of their trunk length , resulting in a lower backrest height . When propelling the chair on a wheelchair accessible treadmill at bot h a level and at a slope of 3° the lo w er backrest height allowed for a greater range of motion with increase d stroke angle, in addition to decreasing cadence. These kinematic parameters were measured using a motion analysis system in addition to using the Smart Wheel system for both wheels [40] . Clinicians often balance. Doing so often also improve s [38] . One reason for individually configuring this angle is to attempt to


18 buttock. Having recruited 22 male healthy subjects with an average age of 27.2 and 22 male s with s pinal cord injury (SPI) with an average age of 32.5 , Park and Jang demonstrated that adjusting the seat to backrest angle between 90° and 130° varied the pre ssures experienced at the ischial tuberosity (IT) as well as sacrococcygeal (SC) area. Park and Jang used a pressure mat with 1,558 sensors and took 5 pressure readings at each 10° interval. For each interval studied , an average pressure was then calculate d. The average peak pressure s of the IT area decreased with increased seat to backrest angle for both healthy and SPI subjects [41] . However, seat and backrest angle s also impact ed the biomechanics behind manually propelling wheelchairs. Manuel et al studied 140 patients with Spinal Cord Injury ( SPI ) comparing seat angle to r ate of developm ent of shoulder pain and injury . Conclusions were based on results from n uclear magnetic resonance imaging, measuring shoulder joint range , and the Wheel chair User Shoulder Pain Index T est . Manuel found that individuals whose seat to backres t angle was at 90° developed shoulder pain at a rate of 1.86 times higher than those with a greater backrest angle. Additionally, individuals with the backrest angle of 90° presented 1.73 times the number of structural injuries as seen in the nuclear magne tic resonance imaging than other seat to backrest angles [42] . Using motion capture systems to capture wrist kinematics coupled with EMG and electrogoniometers (to measure joint angles), a study by Wei showed tha t vertical seat position changed temporal and kinematic parameters during propulsion. Eleven individuals (age range 18 to 40) with various disabilities resulting in wheelchair use were recruited. Each user was place in a wheelchair with a seat that was adjustable vertically and horizontally. At each adjustment vertically and horizontally the participant was asked to propel the chair on rollers for ten pushes at a pace set by a metronome. Gloves re corded duration of the recovery and propulsion phases while muscle activity during each phase was recorded via EMG. An electrogoniometer was used to measure wrist flexion/extension. When the seat height was increased the wrist peak flexion angle increased however a lower seat height increased the


19 ROM. Additionally, radial wrist deviation occurred more often from a lower seat position, while a higher seat position resulted in greater ulnar deviation. As the EMG results showed similar levels of muscle contrac tion/activation at all seat positions, Wei concluded that t here was no ideal seat position. Wei further stated that s eat position should be adjusted as the user experiences radial or ulnar pain to change th e radial/ulnar wrist deviations [43] . While the push rim propelled wheelchair is the most commonly used manual wheelchair , other variations of wheelchairs also exist. By c hanging the propulsion method and forces required during propulsion, each wheelchair adaptation attempts to protect users from muscle and/or joint injuries. However, for various reasons that will be discussed, none have gathered significant traction in th e wheelchair community. G eared M anual W heelchairs A new and slightly modified version of the pushrim propelled chair employs a gearing ratio built directly into the wheels. Several commercially available geared system have been developed. One option is The Easy Push by IntelliWheels. Another such commercially available option is the MAGICwheels. This product provides a 2:1 gearing ratio advantage specifically to decrease demands on the shoulder joint by reducing the necessary propulsion forces. The g earing ratio is button activated allowing the user to engage/disengage the gear at any time. The MAGICwheel also provides an automatic hill holder feature to prevent a manual wheelchair from rolling backwards down a hill [44] .


20 An initial study , over a period of twenty week s, found individuals using the MAGICwheels experienced a significan t decrease in shoulder pain . Seventeen full time manual wheelchair users experiencing shoulder pain were given a four week baseline where they continued to user their own chairs and wheels. Immediately f ollowing the baseline, a twenty week/ five month intervention phase began during wh ich the participants exclusively used the MAGICwheel The intervention phase was followed by a four week retention phase in which the participants returned to using their own wheels. During each phase, the participants were given a weekly assessment to determine changes in shoulder pain. Shoulder pain was determined by a score of at least 10 on the Wheelchair Users Shoulder Pain Index (WUSPI). To ensure consistency in environmental fa ctors between participants, part ic ipants were prescreened to ensure each at least daily encountered various challenges in their environment , from hills or uneven ground. By week two of the intervention phase there was a significant decrease in shoulder pai n as compared to the baseline . During the retention phase , results from the WUSPI showed a trend toward increasing shoulder pain . This trend was attributed to the return to their original wheels [45] . Figure 5 : Example of geared wheel, MAGICwheel


21 A kinematic study by Jahanian , et al used a motion tracking system from Vicon to test differences between a geared wheelchair a n d a standard wheelchair as the parti cipants propelled down a twenty foot section of tiled level floor. I n addition , each participant prope lled a wheelchair on an ADA wheelchair ramp . Six abled body participants with no wheelchair experience between the ages of 20 23 were enrolled in the study. When the kinematic data was analyzed velocity, push time, and stroke distance were decreased when using a geared wheelchair. Furthermore , the shoulder joint kinematics did not differ between regular and geared wheels . No difference s in kinematics w ere detected because the method of applying force to the wheels remained uncha nged. One main disadvantage to the geared wheel system was an increase in strokes needed to accomplish the same tasks. The authors caution ed that the increase in number of strokes needed for each task may offset other advantages [46] . A study by Howarth et al recruited thirteen abled body young adults to use a geared sy stem and a standard whee lchair to compare muscular activity/demand differences between the systems. Each participant had surface electrodes attached to seven different muscles on the right shoulder and arm. Then the participant rolled a chair up a fixe d ramp of 2.44 meters in len gth, while EMG data was collected from the electrodes . From the filtered and amplified EMG data , Howarth showed a geared system decreased peak shoulder muscular demand . However , the integrated EMG data demonstrated an increase in total muscular effort when employing a geared system . When using the geared system, individuals required a greater number of strokes to cover the same distance. This increase in strokes required a greater amount of muscle use, resulting in more muscular effort [47] . R everse /P ull M anual W heelchairs While forward propulsion methods are the most common , some manual wheelchairs are configured to allow the user to use a rowing (backward) motion instead. One such example is the geared rowing wheels called Rowheels [48] .


22 Figure 6 : Two Versions of the Rowheels [49] I n addition to providing a 3:1 gearing ratio , these wheels also require the use of a rowing motion in place of the normal pushing motion. The rowing motion is expected to increase the activity of larger muscle gro ups not normally used during forward pushrim motion. To create forward motion from a pulling action, t he Rowheels employ a planetary gear system that translates a pulling/reverse motion on the rim into a forward motion of the wheelchair. Using surface EMG to quantify muscle activity combined with motion capture, ten able body participants (ages 20 to 27) propelled both a forward and reverse chair on a roller system. Abled body participants were chosen to prevent bias from experience using forward motion met hods . The study reported that the reverse motion resulted in an increased velocity over the same participants using a forward motion . The study also reported an overall increase in muscle activity and thereby metabolic demand during reverse wheelchair prop ulsion [48] . A similar study involving ten able body participants using a reverse wheeling technique found a significant increase in oxygen uptake , ventilation , lactate build up, and heart rate as compared to using a


23 forward wheeling technique. Reverse wheeling did allow for a decrea se in st roke rate when propelling the chairs at the same velocity. While there was a difference in physiological conditions, individual power output was not significant ly differen t between forward and reverse wheeling . The authors of both studies suggested that a combination of forward and reverse methods could be employed to reduce overuse syndromes from only using forward wheelchair propulsion techniques [50] . Both studies would have made stronger cases for the necessity of using a reverse propulsion wheelchair should full time manual wheelchair users have been studied. It has been previously shown that novice wheelchair users often use different push patterns than expert user s . Varied push patterns involve different muscle activations. In a ddition , large kinematic differences have been seen between able bodied and expert paraplegic subjects [37] [51] . C rank Propelled Wheelchairs ( H andcycle W heelchairs) A crank propelled wheelchair incorporates hand pedals that are connected to a wheel by chains and gears . The user can then use a peddling motion with their arms to move the chair. These types of Figure 7 : Handcycle Wheelchair [23]


24 chairs have been shown to be more efficient than the more common pushrim propelled chairs. A study by Dallmeijer examining use of both pushrim propelled chair and a handcycle chair found a significantly lower oxygen uptake tests using a handcycle chair . Th ese differences existed for both the ten abled bodied participants as well as the nine participants with paraplegia. Physiological data was collected while the participants propelled a handcycle and a wheelchair on a wheelchair accessible treadmill. Each test consisted of two different four minute submaximal exercise bouts at set energy output s . During the testing , the subjects wore a computer ized gas analyzing system along with a heart rate monitor. The captured physiological data indicated handcycles place d less stress on the cardiovascular system as compared to the pushrim wheelchairs [52 ] . Another advantage of a handcycle over a standard pushrim wheelchair is th e ability to change gear ratios (akin to the geared wheels) , allowing for greater velocity and reduction of frequency of movements . While studying the kinematics of eight abled bodied individuals on adjustable sport handcycles, i t w as shown that values of angular accelerations in the shoulder during flexion/extension use are in the same range as those in push rim chairs . As the crank propelled chair does not decrease accelerations or greatly decrease range of motion, it is likely that crank propelled wheelchairs do not provide any benefit in terms of lowering chances of developing joint pain. The kinematic data was collected while each participant performed a n ergocycle with a resistance mimicking an uphill climb [53] .


25 L ever P ropelled W heel chairs Another human powered chair uses levers with ratchetting mechanisms that can than apply the s hand motion to the rear wheels . The added length from the levers provides significant mechanical advantage to the user. A study of ten abled bodied male participants compared physiological parameters and energy expenditure between lever and pushrim propelled wheelchairs. Each chair was propelled on a treadmill while participant ventilation, heart rate, and power output were monitored. Based on a t test to compare between wheelchair options, individuals using the lever propelled chair had decreased oxygen uptake and heart rate. Furthermore , the mechanical efficiency of the lever propelled chair was significantly higher than t he efficiency of the pushrim chair, especially during hill climbs [54] . Figure 8 :The Wijit Lever Wheelchair [25]


26 P ower A ssist W heelchair A ddons When manual wheelchair propulsion starts to become overly burdensome or painful, a power assist drive can provide alternative option to switching entirely to a powerchair . Power assist drives are motor drive units that can be attached to a manual wheelchair and can be engaged only when additional help is needed . By providing additional propulsion help, p ower assist a dd ons may provide an alternative mode of transportation that preserves arm function and allows the user to remain physically fit. When analyzing heart rate and course completion rate of fifteen individuals with tetraplegia , power assist made a significant difference. Power assisted propulsion methods reduced cardiovascular (heart rate) and respiratory strain over carpet, dimple strips, up a ramp, and over a curb cut . Participants rated each task as being significantly easier [55] . User questionnaires have been used to follow twenty full time manual wheelchair users over a course of sixteen weeks while using a power assist device, and then a four week period without a power assist. During this time , most participants reported feeling less fatigued from many daily act ivities. Eighteen of the participants reported that access and use was far less limited by terrain [56] . Other questionnaires given to individuals using power assist devices found that p articipants preferred a simple hand rim wheelchair for tasks requiring greater control, and like d power assist chairs for tasks requiring greater force (up hills) [57] . Some work has been done to determine kinematic differences in terms of range of motion and stroke frequency between manual and power assisted methods. One study examined kinematic data from ten full time manual wheelchair users at five different speeds with and without power assist added to their personal chair. Once the participant reached a steady state the last thirty seconds of each test was recorded with motion capture. D ata of ten successive strokes in the last thirty seconds was averaged . The range of motion for shoulder flexion/extension and wrist ulnar/radial deviation all decreased with power assist while stroke frequency remained the same. Due to the decrease in range of


27 motion, the authors predicted that power assist should decrease the likelihood of developing joint pain/injury [58] . M anual W heelchair P ropulsion Manual wheelchair push patterns are broken into two main segments : the propulsion phase, and the recovery phase. The propulsion phase occurs while force is being applied to the pushrim of the wheelchair. Recovery phase is the duration of time that forces on the pushrim are at baseline until the next application of user supplied force , which b egins the propulsion phase . The entire duration of both phases is known as the cycle time [13] . Push patterns can be further divided by the motions made during the different phase of propulsion. Specifically, the push patterns are divided based on motions during the recovery phase. Figure 9 : Four Main Push Patterns Recovery Propulsion Propulsion Propulsion Propulsion Recovery Recovery Recovery


28 Several examples of different upper limb kinematic methods are demonstrated in Figure 9 [15] . travel along the pushrim. However, d uring the recovery phase the use ng any path back to the user defined starting position. Once contact with the pushrim is reestablished , propulsion phase starts. While many different stroke patterns can be used, t he four main stroke patterns are as follows Semicircular (SC) shown in the t op left, SLOP (single loop over propulsion) pattern in the top right, DLOP (double loop over propulsion) pattern in the bottom left, ARC in t he bottom right. The bars on the left and right mark the locations of the start of propulsion and recovery [15] . The SC pattern is characterized by the hands , specifically the 3 rd metacarpal joint on the hand , falling below the pushrim during the recovery phase. The SLOP pattern is characterized by the hand rising above the push rim during the recovery phase. The DLOP pattern is characterized by the hand rising and then falling below the pushrim of the chair during recovery phase. The Arc stroke pattern is characterized by the hand staying as close to th e pushrim as possible during recovery phase [15] . Multiple studies have collected data on the most common types of stroke patterns used, with inconsistent results [15] , reported the most common pattern was SLOP at 45% of manual wheelchair users, DLOP at 25%, semicircular at 16%, and arcing as the least common at 14%. Of the 38 individuals with SPI i n the study, 58% used the same push pattern f o r their both dominate and nondominated arm. The specific propulsion patterns were determined by analyzing data collected from motion analysis systems tracking a reflective marker attached to the third Metacarpa lphalangeal joint (MPJ) on both hands. On the other hand, a study by Richter reported finding s of 60% of manual wheelchair users employing the DLOP pattern, 24% used SLOP, 8% used arcing, and 8% SC [59] . Richter states that t he


29 reason s for this discrepancy is unknown, and that the differences may be attributed to injury type, low statistical power, or differences in participant recruiting location s . Richter analyzed data from a motion analysis system tra cking a marker on the third MPJ , following Boningers method , for a total of 25 individuals all who experienced/primarily used manual wheelchairs as their mode of transportation. As mentioned, i njury type resulting in wheelchair use was different between th included individuals with spinal cord injuries as well as spina bifida and spinal lipoma, where Boninger only examined individuals with spinal cord injuries . While these four patterns are considered to be the normal patterns for MWC users, there is some ambiguity between the different patterns. Patterns that exhibited characteristics of both ARC and DLOP, or even ARC and SLOP, are often classified differently b y trained observers. One such example of a difficult pattern to classify occurs when an MWC user flicks their wrist out and up on the release but then travels back along the push rim. This creates characteristics of both the DLOP and the ARC pattern. Autho rs who have discovered these mixed patterns suggest that further studies should be conducted to really understand how the classifications should be made [16] . Each propulsion pattern will result in varied biomechanical stresses placed on the upper limb joints. Any differences in the kinematic parameters during propulsion are all possible factors in the development of pain. For example , range of motion during push rim propulsion is often studied because it is thought that an individual is most likely to be injured while being externally loaded when at the extremes of their own normal physiological range [60] . B iomechanical A nalysis O f S troke P atterns Studying stroke patterns and their differences requires collection of many different propulsion variables. Examples of such are listed below: C adence, the rate of stroke patter n use


30 m [14] during propulsion phase [45] Braking moment, the maximum negative moment applied to the handrim [45] Impact , the maximum rate of increase of applied force to the pushrim [14] Due to the high prevalence of reported pain from m anual wheelchair users, a significant amount of effort has been placed in determining best practices for wheelchair propulsion . Most studies performed focus on understanding range of motion (RoM) during each push pattern, the peak accelerations/joint force s created by the specific movements, and levels of muscle activ ation . It is believed by many that maximal joint RoM is considered a risk factor for joint injury and pain [60] [61] [62] . Using motion capture to track the kinematics of seven MWU all wheelchair athletes, Shimada et al determined that at lower speeds SC pattern resulted in lower elbow flexion and extension accelerations , as compared to other three push patterns . In addition, accelerations during shoulder flexion and extension were also lower than accelerations seen in the other patterns. However, those who utilized the DLOP pattern had the lowest shoulder abduction/adduction accelerations. These results were obtained from each participant propelled a standardized wheelchair on a wheelchair dynamom eter in 3 minute intervals. Another major finding was that the semicircular pattern allowed the user to spend more time in propulsion phase and less time in recovery phase. Although the SC pattern resulted in the greatest range of motion (ROM), the motion was still with in normal ROM limits . While recognizing the small study size, t he authors concluded the SC pattern resulted in a more biomechanically efficient wheelchair propulsion when compared to the other three patterns [13] .


31 Kwarciak et al recruited twenty five full time individuals and asked each to propel their wheelchair on a treadmill using all four main push patterns. During each push pattern test, Kwarciak collected EMG data as well as cadence, peak force, contact angle, braking moment and impact from an instrumented wheel (OptiPush) . Based on the EMG and integrated EMG data, the DLOP resulted in the lowest muscle activity in the observed muscles during manual wheelchair propu lsion on a flat surface . Furthermore, the SC pattern resulted in the lowest peak force and impact. He concluded the study by stating DLOP and SC patterns were the best for use on a flat surface as they had the lowest cadences and smallest braking moments w hich most closely following Clinical Practice Guidelines [59] . Unfortunately , it is impossible for a MWC user to always travel slowly on smooth level ground . The type of terrain being traversed influences pushrim propulsion techniques. Slowik recruited 170 individuals with parplegia free of shoulder pain and asked them to propel their own chairs on a n ergometer. Each subject was asked to propel their chair in three different scenarios, at their own pace, at their personal highest speed , and at a simulated incline. The Smart Wheel was used to collect wheel contact parameters. As per the norm, kinematic data was collected with a motion analysis system, with the 3 rd MPJ being tracked to determine stroke pattern type . I ndividuals familiar with the standard four propulsion methods then examined the kinematic graph of each stroke and then determined the classification. Slowik determined as propulsion speed increased, fewer individuals used under rim hand patterns such as DLOP or SC. Additionally as the incline grade increased , the participants adjusted their method and began keeping their hands closer to the handrim throughout the push cycle. This resulted in a greater use of the ARC/arcing pattern. This is logical as the user must keep their hand near the handrim to prevent the wheel chair from rolling backwards . By keeping their hands near the handrim the distance their hands traveled was decreased, resulting in a fas ter cadence to compensate for the hill . As a user transitions from pushing across flat ground to a hill the parameters of cadence, force , and contact angle all increase d [4] .


32 Slowik also began to explore a quantitative method of classifying push/stroke patterns. Using two custom parameters net radial thickness (NRT) and total radial thickness(TRT), he was able to create a two dimen sional plot with each axis being one of the custom parameters. This plot was subdivided into four regions that each corresponded to the four commonly observed and studied stroke patterns. NRT was essentially a calculation of how clockwise/counter clockwise pattern was. This allowed for differentiation between over rim (e.g. SL) patterns from under rim patterns (e.g. SC). TRT determined how close the hand stayed to the pushrim during the recovery phase. Small TRT values correspon ded to patterns in which the hand remained close to the handrim (e.g. ARC) while lar ge TRT values represented patterns in which the hand travels farther away from the hand r im . After setting basic thresholds, this proposed system agreed 90% with a rater/per son visually identifying each pattern [4] . F actors R elating T o I njury Besides reported pain levels or medical imaging techniques, n erve conduction studies are commonly used to determine/evaluate the function of motor or sensory neurons via electrical conduction. The individual undergoing this test normally have needle electrodes placed into the location of interest at a predefined distance apart. The subject is then e lectrically stimulated and the activity of one of the innervated muscles is measured. Time delay of the signal (latency) and signal amplitude are among the values recorded. An i ncreased latency is normally associated with pathology [63] . Using a nerve conduction study, Boninger showed median nerve latency increased with increased subject weight . Boninger performed a study with thirty four randomly recruite d full time manual wheelchair users with paraplegia. Each subject underwent a bilateral median nerve conduction study while propelling their own wheelchair outfitted bilateral with Smart Wheels . The Smart Wheels captured forces and moments for all three global reference directions. From the force data acquired regression


33 line s w ere created for height and weight against sensory amplitude and mean median latency. As predicted , nerve latency increased as subject wei ght increased . Subject height significantly corresponded to an increase in median nerve amplitude and weight are two of the most common factors relating to injury/pain [63] . He also demonstrated that an increase in strenuous conditions placed upon the user resulted in greater median nerve function even when normalized for weight. To normalize for weight , Boni nger More strenuous conditions (beyond those created by extra weight) evoked more effort. This increased effort can be the result of uneven terrain or even the rolling resistance of the wheelcha ir. The rolling resistance of the wheel chair is a f unction of many variable s including user weight, low tire pressure, malalignment of wheels, poor weight balancing in the chair . These variables and more result in greater forces on the front castors and e ven deterioration of the bearings in the chair. The authors noted that even though there is a connection between strenuous conditions and injury, individuals competing in wheelchair sports do not demonstrate an increased propensity towards injury . The impl ications being that additional factors may play a significant role in the development of upper limb injury [63] . High repetition of motions or tasks is a common factor in the development of joint pain [64] . Moti ons that are repeated with great frequency or for prolonged period s of time can cause muscle and tendon strain . These strains are caused by an inadequate recovery time between each applied muscular force . The lack of recovery time can lead to overuse which can cause inflammation of the joints as well as increased pressure onto the nerves . If stressed enough , the repeated loading may even lead to tears in the muscles themselves. The Iowa State University website on ergonomics states that any task involving t he shoulders that has a repetition rate greater than 2.5 per minute is at high level of risk especially when combined with high static load, speed, or extreme posture [64] (all of which may occur during manual wheelchair propulsion). Furthermore, it has been estimated that the average number of


34 wheelchair strokes per day is around 2,000 to 3,000 [65] . The clinical practice guidelines for preserving upper limb function after a spinal cord injury , state that effort should be taken to minimize the frequency of repetitive upper limb tasks, such as during manual wheelchair propulsion, as strokes can occur approximately once per second . This rate far exceeds what many studies consider a frequent task . These guidelines were written by a panel of experts in spinal cord injury care, based on their extensive li terature review of spinal cord injury [66] . While most studies specifically concerning overuse injuries in manual wheelchair users have a relatively small number of subjects, work place injury studies provide additional information into overuse sy ndrome. For example, studies have been performed examining over 4,000 workers at almost 20 workplaces all dealing with tasks that are considered repetitive . Findings showed a significant increase (approximately 40%) in shoulder tendinitis when the task inv olved required a minimum of 10% of the maximal voluntary contraction force. Physicians diagnosed findings of tendinitis according to a defined protocol . Findings were based on reported pain during certain movements and activity impairment scales. The force requirements were defined as a light force is < 10%, a somewhat hard (10 29%), hard (30 49%), very hard (50 79%) of the individual s maximal voluntary contraction force [67] . This defined scale of muscle activation and the inc rease in tendinitis directly applied to manual wheelchair propulsion. EMG levels recorded during manual wheelchair propulsion often far exceed the 10% value of maximal voluntary contraction [47] [59] [48] . As discussed earlier, wheelchair parameters are a factor in kinematic characteristics of wheelchair propulsion and therefore are factors in the risk of developing joint pain/injury . Propulsion technique s , wheelchair configuration, and the relationship between them must be carefully considered. The clinical recommendations to preserve upper limb function in spinal cord injury patients suggests actions should be taken to help limit the number of extreme positions experienced by a wheelchair ly, the recommendations warn that any potentially injurious motions should be


35 avoided to protec t the shoulder by limiting rotation and abduction at the shoulder [66] . T hese goals can be accomplished through monitoring propulsion technique and ensuring proper wheelchair configuration [13] [59] . It ha s also been shown that there is a strong correlation between amount of time in a wheelchair to pain and injury. Individuals reporting shoulder pain ha ve spent an average of 19 years in a manual wheelchair, while those not reporting shoulder pain spent an a verage of 9.1 years [42] . Drawing conclusions from the studies referenced earlier, t his correlation should be expected considering the development of joint pain/injury is believed to be the result of overloading/repetitive motions. The longer an individual uses a manual wheelchair, the more propulsion strokes the individual will use . Logically, any effects of small irregularities in the propulsion patterns resulting in greater biomechanical stresse s will be multiplied by the repetitive nature of wheelchair propulsion [64] . P athologies F rom Manual Wheelchair Propulsion An additional factor related to joint pain/injury is the magnitude of the forces experienced by the joints. Jennifer Mercer and Michael Boninger u s ed a motion analysis system , as well as the Smart Wheel to collect kinetic data from 33 individuals with spinal cord injuries (SCI). No individuals currently experiencing joint pain were enrolled in the study. All participants enrolled underwent both MRI and a physical exam by a licensed practitioner looking for various signs of shoulder pathologies . The physical exam was focused on signs/symptoms of shoulder injury focused around pain and discomfort based on various movements and rotations. The MRI was performed on the non dominate side to lessen the chances of a work or leisure related activity skewing pathology findings. Kinematic data was collected while e ach participant propelled their chair at two different speeds on a dynamomete r . Strong correlations were found between joint rotations and accele rations and resulting pathologies. G reater posterior, lateral , or extension moment during propulsion resulted in an increased likelihood of having


36 coracoacromial ligament edema (fluid in the tissue) . Individuals experiencing larger lateral forces or larger abduction moments experienced a greater likelihood of coracoacromial ligament thickening. H igher superior forces and internal rotation moments were associated with increased signs of varying shoulder pathologies. Shockingly, out of this entire group only one subject did not present any abnormalities as found by the MRI test. Additionally, i ndividuals who experienced pain or discomfort found during the physical exam tended to be older and weigh more [14] . Inflammation and edema often mark overuse/repetitive use syndrome . Unfortunately, an MRI diagnosis is expensive and not a clinically cost efficient method of an early detection of joint injury. It is highly concerning that none of the subjects had joint pain ; yet almost all still expressed s ome form of patholog y. Boninger was also involved in another stud y involving shoulder pathology and individuals with paraplegia. He recruited twenty eight manual wheelchair users with paraplegia and had each undergo a physical examination for signs of shoulder injury based on pain following specific motions. Again , each individual also underwent an MRI. A trained radiologist then interpreted the images without know ledge of patient reported pain or the results of the physical examination. Based on the physical examination fifteen of the subjects (54%) had at least one abnormality, while only nine of the subjects had previously reported shoulder pain. Only nine of the twenty eight individuals had MRI radiographs that we re completely normal. A linear regression of the data showed that only BMI was related to a finding in the physical examination or MRI [3] . Using non contr ast clinical shoulder MRI scans to document shoulder pathology , Morrow et al discovered that 70% of examined manual wh eelchair users had shoulder tendon tears. Many of the tendon tears found during the MRI scans occurred at the tendon insertion site. In addition , the MRI studies found coracoacromial ligament thickening, and acromioclavicular (AC) joint edema . E very indivi dual enrolled had some level of AC joint degenerative arthrosis. The study showed u p to 6 0 % of MWC users , out of their subject pool of ten , end up having rotator cuff tears in the supraspinatus .


37 significant difference from . Morrow only allowed participants currently experiencing joint pain to be enrolled. It should be noted as well that s even of the [68] . A study using ultrasound found that of 67 full time MWU enrolled , all but one demonstrated s ome sign of shoulder pathology. 47% of the subjects had mild tendinosis in their bicep. 61.2% had signs of supraspinatus impingement , and strikingly 85.0% had some form of cortical irregularity. Only individuals who were MWU and did not have a degenerative disease were enrolled. Most of the participants were recruited from the National Veterans Wheelchair Games , and a ll ultrasounds were performed on the non dominate arm . The study authors determined that as tendinosis became more severe the ultrasonic image showed less greyscale variance, entropy, and contrast. The authors suggest ed this could lead to additional methods of detecting or monitoring the progression of shoulder abnormal pathology. Other findings include a rate of 49.3% having experienced joint p ain in the last month, and a total of 52.2% having signs of pain/discomfort on the non dominate shoulder as found during the physical examination [69] . Slowik study on the influence of wheelchair seat angle to shoulder pain, found , of their 140 subjects , that 50.7% or 71 individuals displayed signs of pathology f ound by the nuclear magnetic resonance measurements . This is a slightly lower reported percentage of pathologies found from other radiographic studies, but this study did also have one of the larg est recruitments. Furthermore, o f the 57 individuals who reported currently experiencing pain , 93% had injuries. Most injuries/pathologies found were tendinitis or tearing in the supraspinatus and/or infraspinatus [4 2] . After considering the high rate of pain and injury from manual wheelchair use something more must be done . A logical next step would be to determin e effective methods of coaching/assisting MWU in the correct push patterns , shortly after using irregular propulsion patterns . Using the push patterns


38 determined to have the smallest range of motions and greatest efficiencies (DLOP, SC) may decrease the rate of injury. This is where human activity recognition algorithms can be implemented to monitor push patterns used. A ctivity R ecognition T echniques Human activity recognition (HAR) is the attempt to use data from wearable sensors or cameras to train computers to determine what type of activity a subject is performin g. The most common tool used in HAR research is the inertial measurement unit (IMU) . An IMU is an electronic device that can measure linear acceleration, angular rotational rate, and often magnetic field strength . Of these, t he most common type of sensor use d for HAR is the triaxial accelerometer . If cleverly analyzed, predictions can be made as to the type of task that an individual is performing [70] [71] . For example, r eal time monitoring of human motion activity has been explored in fall detection [72] , gait pattern and posture analysis [73] . Just as MWC propulsion patterns relate to shoulder injury , walking patterns are associated with level s of impairment and health issues. P atients who had experienced a stroke demo nstrated a mixture of deviations from normal gait patterns [74] . When using an IMU , the linear acceleration sig nal is a linear combination of a ccelerations caused by change in user movement, acceleration due to gravity, and noise intri nsic to the specific IMU in use. The acceleration due to gravity can hold use ful information as it is always present, and as the the axis of action. Gyroscopes also provide this information and several techniques exist that can remove acceleration due to gravity. This includes taking a derivate of the data as gravity is a constant and will simply fall out. High pass filters can also be used to remove the acceleration due to gravity. Noise intrinsic to the specific IMU presents a different difficulty as thi s noise is often random and has a non zero mean error , making it difficult to calculate a precise location in space from only accelerometer data [75] .


39 Having information pertaining to the acceleration and angular rotation of a body solves only part of the challenge of HA R. The data from specific tasks must be analyzed and compared to the data from other tasks in such a way that differences can be detected. Unfortunately, for many tasks this requires rapidly collecting and shifting through a large amount of highly non line ar data, where no one to one direct mapping function exists. To compensate for this lack of a one to one relationship , various machine learning techniques have been designed specifically for this task. Previous studies have demonstrated that machine learn ing techniques are effective at determining many different types of activities . These determinations are made from data collected by sensors placed on various locations on the user. Depending on the method involve d a t hree step process occurs . F irst sensor data is collected from a known activity and then feature extraction is performed . S econdly, the . Finally, the classifier is given new data and asked to decide what activity it most closely relates to . Feature extraction is a method of data analysis, from which important features or aspects of the data is pulled out of the data. For example, f eature extraction of data can be broken into three different subcategories. F requency analysis examines the feat ures of the data relating to repeating elements in the data, or feature extraction pulls directly from the time varying signal and include information such as mean, standard deviation, normalcy, excreta, and finally wavelet features also k nown as time freq uency features . In the design of such systems one can consider that most human daily activities result in accelerations with frequencies of 20 HZ or less, giving clear locations for data filtering [76] . When using machine learning techniques , specific features extracted as well as the bin /window size of analyz ed data greatly impacts the accuracy of the classification algorithm. A bin is a set number of data samples that are then feed in successive order into the classifier . These bins can be distinct from each other or contain some overlapping with previous bin s. Bins are used when the start and stop of specific activities are unknown. Some examples of features from the statistical domain (time varying)


40 include: Kurtosis test, Skewness test, mean, standard deviation, interquartile range, histogram, root mean squ are, median absolute deviation, and others. In the temporal domain expected features can include: max/median frequency, cepstral coefficients, power spectrum, fundamental frequency, power bandwidth . When all else fails, increasing the amount of data collec ted and analyzed often will increase classification accuracy [75] . Determining which features to use for classification along with an appropriate window size is dependent on the type of task as well as sensors location and type . According to a review, s ome studies have had success with high classification accuracy from time domain features alone, while others have had success with data from frequency analysis. Depending on the task, the mean and standard deviation alone have been reported to give great er than 90% classification accuracy [76] . Like the vast number of possible features that can b e extracted and used, t here exist many different types of machine learning methods which have been applied to HAR. Machine learning methods are broken into two main categories : supervised methods and unsupervised methods. Supervised training methods requir e the user to provide a set of training data with known values/outcomes, then provides a new set of data with unknown outcomes for the system to make predictions. Unsupervised methods , on the other hand , make classifications entirely based off hidden structures in the data . Examples of supervised techniques include k nearest neighbor, naïve Bayes, support vector machines (SVMs), principle component analysis/ linear discriminant analysis (PCA/LDA). Unsupervised models include Gaussians mixture models (G MM) and hidden Markov model (HMM) [77] . One specific study attempted to use fuzzy clustering to find di fferences in the data of various push patterns. Kinematic data was collected using a motion capture (optoelectronic) system. From the kinematic data, two specific parameters were calculated as functions of the thickness of the motion contour line and the t opological aspect of the motion contour line. Fourteen elderly individuals were


41 consented and propelled a custom chair at nine different configurations between the seat to backrest angle and system tilt angle. Pushrim force data was also collected using a Smart Wheel system. Four different patterns were found in the data. Of which three would likely be classified as the arcing pattern, by a rater system, each of different angular lengths. Thresholds were found that could separate the discovered patterns as a basis of their calculated parameters [78] . Any system tha t performed real time analysis would likely need to differentiate between these different sub patterns. Additional studies are needed to determine what other sub classification patterns may exist. Real time analysis of human ambulatory activities has previ ously been studied. Constraints on Real Time analysis include a lack of knowledge of future events, limited data buffering sizes, and limited processing times. Techniques used to analyze data from tri axial acceleration data include low passing the data at 0.25 Hz. Additionally, general classification success has occurred with data sampling rates at 45 Hz [79] . T he R emaining P roblem Even with the number of studies trying to understand the development of joint pain from MWC propulsion, the studies have been limited in many similar ways. Many studies struggle to recruit enough participants to have strong statistical significance . Additionally, nearly every study has been restricted to a one time examination . While few longitudinal studies have been performed , no method currently exists to collect data outside of a clinic where unexpected factors may play a role in joint injury. Furthermore, most studies to this point only focus on wheelchair propulsion resulting in forward motion . For example, k inematics during turning and small corrections during propulsion have yet to b e fully considered. To truly understand the development of joint pain/injury , constant kinematic data from daily wheelchair use is required. This kinematic data should describe the push patterns used and th e relate d


42 environment s being navigated . Additiona lly, the data could describe irregularities in the push patterns . Once these goals have been attained, longitudinal studies will be better equipped to study the development of upper limb joint injury resulting from manual wheelchair propulsion.


43 CHAPTER II I M ETHODOLOGY : Based on the previous discussions , it is predicted that accurate (>90% correct) real time classification of propulsion patterns can be made . Furthermore, it is predicted that individuals experiencing joint pain will use a greater number of ARC and SLOP patterns than individuals not experiencing joint pain. H ypothesis 1 : Using a Custom Activity Tracker (CAT), classification of four (DLOP, SLOP, SC, ARC) propulsion patterns will be identified by the CAT with 90% accuracy. To test this h ypothesis two different phases are required . In the first phase , a custom wrist worn activity tracker (CAT) was designed and built to collect IMU data. T he second phase employ ed several different machine learning algorithms designed to attempt to classify an expert manual wheelchair individual push patterns . Specific Aim 1 .1 : Design and build a custom activity tracker system (akin to a Fitbit) capable of collecting kinematic data (linear acceleration, angular rate of change), and then perform classification of each propulsion stroke. Device Design This project focused on the use of IMU data to classify each propulsion cycle as a SC, SLOP, DLOP, or ARC pattern. The specific IMU data (directly pertaining to each propulsion stroke) collected is listed below: Linear Acceleration Angular Rotation Rate Magnetic Field Strength


44 The cu stom tracker was designed to be worn at the wrist with another unit worn just above the wearer s elbow , a s shown in Figure 10 below. To lower the chances of the CAT interfering with regular propulsion, the trackers were designed to be small and lightweight. Specifically, t he custom tracker was designed to be approximately the same size as commercially available activity trackers. Recognizing the impact of terrain on propulsion pattern strategy, the system was designed to include a small unit that collected wheelchair specific kinematics. The specific kinematic data collected is listed below. Wheelchair V elocity Wheelchair S urface I ncline Wheelchair Accelerations Figure 10 : Data collection Locations


45 Algorithm Design and Real Time Classification: Several different machine learning methods were implemented and tested for their ability to make accurate predictions. The algorithms were trained on features extracted from digitally filtered IMU data collected from the custom tracker. The t hree specific algorithms considered in this study along with brief explanations are listed below. Random Forest A Random Forest algorithm works by generating a succession of decision trees each attempting to minimize the number of different classes in each leaf of the decision tree. The entire set of trees is then a sked to classify new data and the most commonly returned answer is considered the actual class. Artificial Neural Network Artificial Neural Networks were designed to mimic the function of neurons. Each input is combined with the weights assigned to each neuron, through several layers until a final output is given. Each step hopefully getting closer to the correct output. A simple two input with one output neural net is shown below. Hidden Layer 1: Inputs: Weights 1: Weights 2: Output: Figure 11 : Simple 1 Layer 1 Output Neural Net


46 Support Vector Machine A S upport V ector M achine attempts to find a hyperplane that successfully separate s the different classes with the largest margin. The a ssociated support vectors assis t in determining classifications for data points closest to the margin . One support vector machine is only capable of separat ing two classes . To separate multiple classes, m ultiple support vectors machines are then used to separate all available classes. A ll data analysis in this thesis was performed using MATLAB . As such, e ach of the three machine learning methods listed w ere trained and tested using MATLAB functions . Algorithm success was determined using IMU data not included in the initial training data set. This testing data contained known classifications which were compared to the algorithm predicted classifications resulting in a percent accuracy. Specific Aim 1 . 2 : Validate the classification algorithm using data collected from a fulltime manual whe elchair user. After Institutional Review Board review and approval of Protocol submission #18 0115 was obtained, a full time manual wheelchair user was recruited to collect the data necessary to train each classification method . This individual was selecte d as a basis of the criteria listed below. Inclusion Criteria: 1. Full time manual wheelchair user (main mode of transportation in and out of the house). 2. Used a wheelchair longer than 3 years. 3. No significant motor disabilities in upper limbs or hands. 4. Between the ages of 18 and 65 5. Capable of pushing a wheelchair continuously for at least 5 minutes


47 Exclusion Criteria: 1. Uses other means of transportation besides a manual wheelchair a significant portion of the day. 2. Used a wheelchair for less than 3 years. 3. Has si gnificant motor disabilities in upper limbs of hands. 4. Younger than 18 and older than 65. 5. Unable to push a wheelchair continuously for at least 5 minutes. 6. Pregnant or expecting to be pregnant during the duration of participation in this study Table 1 documents the process followed for the collection of the initial training data. Table 1 : Validating classification Protocol Consent and Eligibility Time: 30 60 minutes Consenting and eligibility criteria screening procedures; obtain photo/video release. Eligibility will be confirmed. The participant will be asked to read, ask questions, sign and/or fill out the documents listed below. Forms and Tools: Consent and HIPAA Authorization Form Photo, Video, and Sound Recording Release Demographics and Health History Phase 1: Training Phase Time: 1 hour The expert manual wheel chair user will be asked to familiarize himself with using the 4 common push patterns. Then he/she will be asked to continuously use each type of push pattern for 2 5 minutes or until a minimum of 100 strokes of each of the 4 types of push pattern have been used. The user will be given the opportunity to rest for several minutes , in between each set. Acceleration , Gyroscopic , magnetometer data 100 DLOP 100 SLOP 100 ARC 100 SC Video Data Forms and Tools: Video Recording Custom Activity Tracker System The collected IMU data was then imported into MATLAB for analysis. The raw data was band passed between 0.5 Hz and 6 Hz. Then the individual push patterns were isolated and feature extraction


48 was performed on each push pattern. Data from 95 individual pushes of each propulsion method w as used as the training data while 25 pushes of each classification were withheld as testing data. This training data was then used to train each classifier. The testing data was then used to determine the performance of the trained algorithms. Performance was measured by the percent of testing cycles correctly classified. Various graphs were generated to show the effect of specific algorithm feat ures on the overall accuracy of each algorithm. Additionally, the effects of PCA on each algorithm was tested to determine the minimum necessary number of principle components while still ensuring optimal accuracy. H ypothesis 2: Patients who use a greater number of arcing or single loop propulsion patterns as compared to subjects who use a greater number of semi circular or DLOP propulsion patterns will report significantly higher levels of upper limb joint pain. This study was conducted following protocol #18 0115 reviewed and approved by COMIRB. Any participants were voluntarily recruited and provided informed consent before beginning of any of the activities. An outline of the protocol followed is shown below in Table 2 .


49 Table 2 : Simulated Environment Testing Protocol Inclusion and exclusion criteria for participants was the same as the criteria for the initial training data collection. Phase 2: Pilot Study: Specific Aim 1 : Controlled Environment Time: 1 2 hours Full time manual wheelchair users selected via process described above, will be asked to wear the custom activity tracker, and navigate through a controlled environment simulating normal obstacles encountered in Acceleration and Gyroscopic data Classifications Wheelchair Velocity and Angle Forms and Tools Custom Activity Tracker Specific Aim 2: Uncontrolled Environment Time: >10 hours Full time manual wheelchair users selected via process described above, will be asked to wear the custom activity tracker, and go about a minimum of 10 hours of normal daily life. At the end, the custom activity tracker will be returned and acceleration, g yroscopic, and wheelchair velocity data collected by the activity tracker , will be retrieved, and stored per the data storage plan described later in the document. Only data from the first 24 hours or less of use will be included in the study. Acceleration and Gyroscopic data Classifications Wheelchair Velocity and Angle Forms and Tools Custom Activity Tracker


50 Specific Aim 2. 1 : Using the activity tracker, compare propulsion pattern differences between a population of up to 20 MWU completing a simulated daily living environment. While the protocol was approved to allow up to 20 MWU, only t hree full time manual wheelchair users we re recruited and wore the custom tracker while they navigate d through a simulated area of daily living . The participants started by propelling over a small ADA ramp across a section of tile before turning right and crossing a section of carpet. Then the pa rticipants took a left turn and went up a small ramp twice as ste e p as the ADA ramp. To finish , each participant crossed another tile section before taking a right turn and propelling over another open section of tile. Each participant wore the custom acti vity tracker on their right hand for consistenc y between participants. Additionally, a camera phone was mounted to each of the participants chairs such that the right hand and arm could be recorded during the process. Each participant completed the course a total of 3 times. The data was then imported into MATLAB and manually synchronized with the video data. Using the video data as the guide, the individual propulsion cycles were isolated from the data and underwent feature extraction . The se extracted features w ere fed into the classifiers and the predicted classifications were compared to the classifications assigned during video review. Specific Aim 2 .2 : Using the activity tracker compare, at minimum, full day use (10 hours) of kinematic differences between the before mentioned group in their normal routines. Due t o poor performance on classifying data not collected from the individual who provided the training data, no data analysis was performed on the IMU data collected during regular manual wheelc hair use . Data was collected following the approved COMIRB protocol but not classified or


51 analyzed . Any results returned by the classifiers would have been untrustworthy. This is discussed further in the discussions section.


52 CHAPTER IV BUILD AND DESIGN H ypothesis 1: Using a Custom Activity Tracker system (CAT), Classifications of four propulsion patterns (DLOP, SLOP, SC, ARC) will be identified by the CAT with 90% accuracy. Specific Aim 1 .1 the Design: Design and build a custom activity tracker system (ak in to a Fitbit) capable of collecting kinematic data (linear acceleration and angular rate of change), and the perform classification of each propulsion stroke. The Custom Activity Tracker system required the interaction of multiple subsystems . The CAT sys tem is broken up into the upper limb data collection system, the wheelchair data collection system, and the classification algorithm. Full System Design : Wrist CAT Elbow CAT Wheelchair Data Collection System MATLAB Gauss Magnetic Chair Mount User Motion Wheelchair Velocity Classifications Figure 12 : Full System Flow Chart


53 Figure 12 above depicts the process flow diagram followed for collecting and performing data analysi s . IMU data could be collected from both the elbow and wrist (upper limb data collection system) and then sent via a USB connection to a computer running MATLAB. Simultaneously, the wheelchair velocity as well as accelerations and slope change were collected by the wheelchair data collection system. All data was loaded into MATLAB where the classification s were performed. Upper limb data collection system: Propulsion patterns used to maneuver manual wheelchairs rely primarily on the glenohumeral joint, as well as the elbow joint to move the location of the hands in space [14] . As such , two forms of the activity tracker were built. One that is capable of monitoring kinematics from both the elbow and wrist, as well as one that only monitors the wrist. This set up treats the individuals arm as a rigid body two segment system. This allows all arm motions during manual wheelchair propulsion to be captured . Figure 13 : Arm Segments


54 In Figure 13 above, the yellow balls represent the joints where independent motions can be generated. It is possible to move the hand by flexing or extending the elbow without rotation at the shoulder joint. Additionally, the arm can be rotated about the shoulder while locking the elbow. This makes it possible that the propulsion patterns will have differences at both the wrist and on the upper arm. The blue rectangles depict the locations for the CAT where the accelerations and angular rates of rotation are measu red. Flow Chart CAT : Information was collected from the onboard LSM9DS0 chip as well as from the hall effect sensor and fed into the Adalogger processor. At this iteration of the CAT, no data processing occurs onboard the device . E ach data sample collected was immediately saved to the SD card in a comma separated variable text file. Software can be uploaded and updated via the microUSB port on the device. Additionally, the data from the microSD card c ould be streamed across the serial connection of the USB to allow the data to be copied to a computer where MATLAB c ould be used to process the data. SPI LSM9DS0 (Accelerometer, Gyro, Magnometer) Adafruit Feather M0 Adalogger (microprocessor) Battery SD Card Hall Effect Sensor Power Switch 100 mA USB 5V Serial User Motion Gauss Figure 14 : CAT Flow Chart 100 mA


55 Main Board: The main board running the CAT is the Adafruit Feather M0 Adalogger. It is a complete board that is 22.8 mm x 51.6 mm x 8 mm in size and weighs 5.3 grams . At its core is the ATSAMD21G18 ARM Cortex M0 processor. This processor is clocked at 48 MHz and operates at 3.3V logic. Additionally, the chip c ould handle I2C as well as SPI communication methods. The board also c ame with a built in M icroSD card holder that use d SPI to write data to an inserted MicroSD card. Adafruit ships the board with a USB bootloader already loaded on the system to all it to be easily programmed from a computer using the Arduino i ntegrated development environment (IDE). This allow ed for a quicker development and access to a significant number of libraries already created to interface with multiple hardware types. The Adafruit Fe ather M0 Adalogger also contain ed a built in 100 mA Lithium Polymer battery charger . The charger automatically starts when the device was connected to a power source via a USB cable. Wh ile a USB cable was connected , the board was powered through the cable rather than by battery . The system then switch ed back to battery power when unplugged. The average current draw of the system was approximately 10 mA with peaks as high as 30 mA . These peaks are caused when large amounts of data are written to the SD card . T he board manufacturer measured and reported these values. Based on testing data, when data is being streamed via a USB cable to a computer, the max sampling rate of the board was approximately 155 Hz, with an IMU measurement being taken ever y 6.3 ms. W hen data was being saved to the MicroSD card the data rate is decreased. The sampling rate then bec ame approximately 48 Hz . Human activity is generally considered to be below 20 Hz . E xcept in extreme and infrequent situations , propulsion rate will rarely exceed a max of 3 4 H z . As such , this sampling rate is above the Nyquist frequency and should prevent aliasing of frequencies. Additionally ,


56 recent stud ies have measur ed the effects of accelerometer sampling rate on classification of human activities . One specific study found that a sampling rate of approximately 45 Hz was sufficient high to prevent significant decrease s in classification accuracies [80] . Sensors: Each tracker w as built using t he LSM9DS0 chip that contains a 3 directional Accelerometer, 3 directional Gyro, and a 3 directional Magnetometer . The sensor can measure linear acceleration with sensitivity at ± 2, ± 4, ± 6, ± 8, ± 16 g forces. Additionally, the sensor measures the magnetic field with sensitivity at ± 2, ± 4, ± 8, ± 12 gauss in each direction. The sensor measures angular rate with sensitivity at ± 245, ± 500, ± 2000 degrees per second (DPS) . The sensor supports an ser ial bus capable of transmitting data at either 100 kHz or at 400 kHz to a microprocessor. The chip can function in a temperature range between 40 C all the way to +85 C. As this system was a battery operated device , normal operating current consumption was very important. The data sheet states that when the chip was in normal mode current consumption of the accelerometer is approximately 350 µ A and the gyroscope has a current consumption of 6.1 µ A. These low operating currents will help increase total b attery life. Based on preliminary propulsion pattern data, the sensitivities in the sensor were set with values of Acceleration at 2G, Magnetic Field at 2 Gauss, Angular Acceleration at 245 DPS. Error Calculations : The sensor was tested to determine an approximate level of random noise in each data stream from the LSM9DS0, as well as determine level of offset based on the expected acceleration due to


57 gravity. For each test the sensor was placed in a specific orientat ion such that the axis being tested was perpendicular to gravity. The sensor was placed in each specific orientation and left for approximately 10 seconds while data was collected. Z: The average value during the test was 9.8054 with a standard devia tion of 0.0220 . The largest range of values during the test was a difference of 0.3 . Additionally, the x axis had an average of 1.5629 and the y axis had an average of 1.2686 X: The average value during the test was 10.0244 with a standard deviation of 0.0173 . The total range of the values was 0.18 . Additionally, the y axis had an average of 1.5416 , and the z axis had an average of 1.3275 . Y: The average values during the test was 9.6170 with a standard deviation of 0.0095 . The range of the values during the test was 0.05 . Additionally, the x axis had an average value of 0.268 , and the z axis had an average value of 1.4029 . Hall Effect Sensor: To assist in determining when an entire propulsion cycle was completed a hall effect sensor was e mbedded into the wrist band so that a magnet placed on the wheelchair c ould be detected. Every time the user passes by the magnet , the user should be in a simila r phase of the propulsion pattern. The hall effect sensor chosen was the AH9250 by Diodes I ncorporated. This sensor operates with a power supply between 2.5V and 5.5V. When operated at 3V the reported average current consumption is 8 µ A. This


58 incredibly lo w current consumption was due to a sleep mode cycle used by the sensor. The sensor wakes up for 150 µ s and then latches the output for 100ms. This sleep cycle fixes t he maximum switching frequency to approximately 10 Hz. Push rim propulsion is often much l ower than 10 Hz and should likely never exceed 4 to 5 Hz. Additionally, the chip comes with a built in pull up resistor, meaning less connections are needed for proper chip operation. While the manufacture recommends connecting capacitors between the outp ut and ground as well as power and ground, initial tests demonstrated sufficient functioning without the recommended capacitors. The chip is omni polar and can will toggle when the magnetic flux perpendicular to its face was approximately 40 Gauss. From fu nctional tests , this means that the hall effect sensor can detect the selected neodymium magnets from approximately 1 inch away , even when the hall effect sensor was encapsulated in silicone. Battery: The selected battery was the PKCell LIPO 552035 350mAh 3.7V. This battery includes built in battery charging and discharging protection circuitry. This means that the battery can shut itself off when it has been discharged to a set point . This helps prevent the battery from being over used and potentially damaging the battery. Additionally , the battery manufacturer reports battery tests in conditions of being crushed by approximately 3000 pounds , being dropped , and being smashed with a 20+ pound weight. These a nd other tests performed demons trate d that the battery should be safe for use in a wearable device. The battery takes up a volume of 36mm x 20mm x 5.6mm and weights 7.9 grams. This sizing was important to ensure the device will be small enough to be worn comfortable and not add additi onal strain to the user.


59 Based on the main board pulling 10 mA, the sensor pulling 360 µ A, and the hall effect sensor pulling approximately 8 µA. The entire system should between 10.37 and at max puls 30.37 mA. The entire system then should be expected to operate for some time between 11.52 hours and 33.75 hours, before the battery will turn off. In either case the system c ould collect data from at least 10 hour of normal wheelchair use. Device Design : Once the specific electronics were picked , they were each modeled in solid works to determine the optimal placement for smallest volume. It was determined that the LSM9DS0 could be attached to the bottom of the A dalogger. The battery could sit directly on top of the board . I n this configuration the Figure 15 : Electronics Dimensioning


60 total system was approximately 13mm tall. The Adalogger was the longest component as well as the widest with a base of 23 mm by 53 mm. For size comparison , the Samsung Gear Fit 2 main board is 50 mm by 23 mm with a height of 10 mm. The Fitbit Ionic has a screen size of 36 mm diagonal. The system was not significantly larger than other custom activity trackers on the market. The connector was re moved from the battery to allow access to the two wires which were then connected to the power switch . The power switch was then connected directly to the battery connector on the board. This allow ed the system to be turned on and off without removing the battery. The downside to this configuration was the switch must be in the on position to charge the battery. The A dalogger c ould detect when a MicroUSB is connected, and switch from battery power to the M icroUSB and begin to charge the battery. Figure 16 Depicts the simple circuit used to toggle power for the CAT. When S1 is on the off position, no electrical connection is made between G1 + and the positive terminal on the Adalogger. As such the device remains off until the switch is toggled. Figure 16 :Battery Circuit


61 Figure 17 above shows the 3 d model used to determine the size and layout of components for optimal spacing. Additionally, it was used to determine tolerances in the mold and expected silicone wall thi cknesses. Figure 18 : Watch Dimensions Figure 17 : 3D Circuit Model


62 Once the circuit was designed and the final dimensions determined, a watch band was designed to safely house the electronics. Two primary functions drove the design. First the watch needed to be adjustable to a wide range of sizes and secondly it needed to be comfortable. The total watch length is 226 mm with an adjustable range of 49mm in steps of 7 mm. According to online sources, a the normal size of small bracelet is 178 mm . A medium size bracelet is at approximately 203 mm with a large siz e of approximately 228 mm [81] . Dragon Skin 10 NV was chosen as the material for the watch band. The D ragon S kin is a silicone rubber that is often used in prosthetics at the skin device interface . Dragon S kin is a two part mixture that was combined in equal volumes/weights and allowed to cure into a solid. This makes a flexible and soft material for a watch band that is non abrasive to the skin. Shown above in Figure 19 is the desired end shape of the watch band. The specific dimensions are called out in Figure 18 . It was designed to fully encase the electronics. The initial prototypes did not have a screen to display information or allow interactions. Figure 19 : 3D Watch Model


63 A negative of the watch was turned into a mold that was constructed using additive manufacturing . On the left end of the mold a buckle was held in place with glue and then covered with silicone during the molding process . Ultimately, the buckle was firmly held in place by the cured silicone. Additionally, the hall effect sensor could be placed into the right section of the mold and encased into the silicone . The mold was designed to be open faced so the bottom of the mold would be open to the air to allow e asy leveling of the uncured silicone . Additionally, the open face created an easy escape for Figure 20 : 3D Silicone Mold Model


64 bubbles trapped in the silicone. The silicone chosen used a non vacuum cure with a cure time of 75 minutes . Wheelchair Velocity and Angle Detector: In addition to the arm worn custom activity tracker, a wheelchair velocity and angle detector was also built. This unit shares several design similarities to the custom activity tracker. The main board for the device was the Adafruit MO Adalogger and use d the LSM9DS0 for determining angle and accelerations. Additionally , a hall effect sensor was mounted to the board so that small magnets placed strategically on the wheel c ould be detected and the wheelchair velocity c ould be determined. Wheelchair velocity was determined by placing three magnets onto screws on the wheelchair wheel. A small hall effect sensor was placed onto the body of the wheelchair in line with the magnets rotation . Every time a magnet passed the hall effect sensor the input to the wheelchair velocity and angle uni t changes. Then the wheel rotational rate could be calculated by finding the elapsed time (dt) in milliseconds since the last magnet was detected. The number of revolutions per minute (RPM) can be found using the equation . The wheel velocity can then be calculated using where is the distance from the center of the wheel to the magnets measured in feet. Unt il another magnet is detected, the wheel speed was Figure 21 : Chair Mounting Unit


65 stored as constant. If the current dt ever exceeds the dt of the previous detection than the wheelchair speed can be calculated at each time stamp using the current dt until a magnet was once again detected, resetting the dt count. embedded in the watch, magnets needed to be placed on the wheelchair sticking out toward the hand rim. The figure below sho ws the digital model along with the magnets in their respective places. One magnet on the back was used to mount the piece to the side of a manual wheelchair. The front with the two magnets side by side was used to give greater magnetic field area for bett er chance of triggering the sensor in the watch band. Data Collection: An expert/experienced manual wheelchair user was recruited and consented following an approved COMIRB # 18 0115 to provide the initial data to train the algorithms. The expert user was provided with some training to refresh what each pattern looked like . Th e participant was also allowed some time to become comfortable using each propulsion pattern. While wearing the custom activity tracker , the expert user used a selected pattern to propel themselves back and forth down a long tiled Front Back Figure 22 : Magnet Attachment and Holder


66 hallway . This was repeated several times until a minimum of 120 of each push pattern was collected. T hese push patterns were then randomly divided into training or testing data sets. Three other participants were consented following the same COMIRB approved proto col. These subjects navigated through a simulated daily activity course involving two ramps, tile, and a section of carpet. Each subject wore the CAT and were filmed while navigating through the course. Algorithms: Kinematic data collected during manual w heelchair propulsion takes a time series format . This means that each cycle contains multiple sets of data taken at various time stamps . No one set of data from any one point in time contains all the information describing the push pattern. Additionally, a s the user was repeatedly pushing the wheelchair , the data collected was cyclical in nature. In order to handle the varying lengths of data describing each propulsion pattern , feature extraction was performed on the data. The data was collected and arrange d such that each time point resulted in either a 1x10 or 1x19 matrix. In the case that data from both the wrist and elbow are being collected, time was shifted to the location and then corresponds to the first Accel: X Elbow, with bein g Accel: Y Elbow and continuing for all 9 axes.


67 Table 3 : Data storage format Sample Accel: X Accel: Y Accel: Z Gyro: X Gyro: Y Gyro: Z Mag: X Mag: Y Mag: Z Time: 1 2 3 4 5 6 7 8 9 10 1 2 3 Each successive data point is placed in the row below such that contains the first point from acceleration in the x direction and contains the second data point from the acceleration in the x direction. Next the raw data was band passed between 0.5 Hz and 7 Hz before being normalized. Normalizing Data: Many methods exist for normalizing data. Normalizing data was an important step because each data type collect ed will have a naturally differing range. In this instance, a cceleration had a value between 19.6 and 19.6, while the gyroscope returned a value betwe en 245 and 245. The importance of the data is not reflected in the magnitude of the range, and as such, all data needed to be brought into the same range. Normalizing the data also balances differences between individuals. The simplest method of normali zing the data is to divide each point by the max recorded value, doing so will result in all values being between 1 and 1. However this may not the best method of normalizing the data.


68 Another method is to normalize by the z score. This is accomplished through the following equation. . Each point has the average of the specific axis subtracted and then divided by the standard deviation of that specific axis. A common me thod of normalizing data is the max/min method. Each point is feed into the equation . This results in a scaling of 0 to 1 for each data point in each IMU axis. This is the method used for the rest of the data presented. After normalizing the data, features for each individual push pattern were extracted. Features E xtracted: Below is a list of the features extracted from the normalized and filtered data a long with the equations used to do so. Average: whe re n = number data points in the cycle and L = the IMU sensor/axis being collected. Absolute Value Average: the absolute value of each point is taken before the average is calculated. Standard Deviation: wh ere n = number of data points in the cycle and L= the IMU sensor/axis being collected. Root Mean Squared (RMS):


69 Kurtosis: Where E(t) represents the expected value of the quantity at t. This represent s the measure of how outlier prone a distribution is. If the data being analyzed has a normal distribution the Kurtosis will have a value of 3. If the system is more outlier prone that the kurtosis is greater than 3 and the kurtosis is less than 3 if the d istribution is less outlier prone than the normal distribution. This equation is valid when the data is not biased. If the data is biased (is a sample rather than a population) the following equation applies. Entropy: This is a value that relates to the amount of information/randomness of a signal. Several different equations can be used that relate to the entropy of a signal. Matlab has a function wentropy() in the wavelet analysis toolbox that calculates the Shannon entropy using the equation . Where s is the signal and is the coefficients of s in an orthonormal basis. Slope Sign Changes: The number of times the slope of the data changes during the completion of one propulsion cycle. This is calculated by comparing the backward difference at each point and check the sign of the slope. Should there be a difference is sign between the previously calculated slope and the new slope, as well a s a sufficiently sized slope, the Slop Sign Change counter is incremented. Median: This is the data value that is the midpoint of the observed data point. Any observed data point has an equal probability of being above or below this value. Once the data i s ordered by value, the central data point is the median. Median Absolute Deviation: This is the median value of the transformed data when the median has been subtracted from every point.


70 Structuring Data: Each feature is extracted and stored in a 1x9 array. With each additional row representing the next cycle. Then a specific number of sensors/ directions can be chosen from each feature to be analyzed. A new array is constructed holding the features and sen sors/directions being analyzed. Table 4 :Data Formatting Example Cycle Average Accel X: Average Accel Y: Standard Deviation Accel X: Standard Deviation Accel Y: RMS Accel X: RMS Accel Y: 1 2 3 4 5 6 1 2 3 I f only the accelerations in the x and the y directions are considered and the features being extracted are the average, standard deviation , and root mean square d of each cycle, the table may look something like the one above. During the training process, one table is built per class being studied. In this case, there is a table for the SC, SLOP, DLOP, and ARC pattern. Then a selected number of the cycles are randomly removed from the training data set . The cycles removed are then used to verify how successful ly the classifier has been trained. This acts as a verification of the algorithm. Finally, Principle C omponent A nalysis (PCA) was performed on the data to transfo rm the data set.


71 Principal Component Analysis (PCA): PCA finds the axes in which there is the greatest variability in the data. PCA accomplishes this through the singular value decomposition (SVD) of the correlation matrix. Single value decomposition results in the equation below. X Correlatio n Matrix U Right Rotation S Stretch/Scaling V Left Rotation The U matrix contains the normalized eigenvectors of . While S is a diagonal matrix containing the eigenvalues of . This decomposition can be accomplished in MATLAB using eithe r the SVD() or the EIG() function , which will return the same results . E IG () can only be used on square matrices, while the SVD () function can be used on any sized matrix. M ATLAB uses the Cholesky factorization method to break down the matrix. The training matrix was constructed by stacking each of individual cycle training matrices . The resulting number of rows represent ed the number of push patterns collected , minus the patterns withheld for testing purposes. The correlation matrix is calculat ed using this training matrix inserted into the equation . is the ave rage of each column in the training matrix. Then single value decomposition is then performed on the covariance matrix to get the eigenvalues (S) and eigenvectors (U). The eigenvalues found correspond to the level of variance created by each axis (feature) . The larger the eigenvalue the greater the separation. PCA can be used to simplify the data being analyzed and remove variables that are dependent on other variables . For example, take two different variables being analyzed, the first x has values of 1,


72 2 , 3, 4 and so on. The second variable has values of 2, 4, 6, 8 and so on, the second variable contributes no useful information as its values are a function of the first variable . As such when PCA is performed, the second variable will be removed from the data because the variance/variability in the data is already accounted for by the first variable All training data as well as all testing data samples can then be transformed by the U matrix resulting in a new dimensional space. If only eigenvecto rs associated with the largest eigenvalues are used to transform the data a decrease in dimensionality occurs. This can be used to remove features that are not as useful in separating the data. Once the initial training data has been transformed into the n ew subspace all data collected in the future as well as the testing data must also be transformed into the new subspace. Finally, these transformed training and testing data sets were given to the Random Forest, Artificial Neural Network, and Support Vect or Machines. Additionally, non transformed data sets were also given to the algorithms to see if PCA could improve the accuracy of the algorithms.


73 CHAPTER V R ESULTS H ypothesis 1: Using a Custom Activity Tracker system (CAT), Classifications of four propulsion patterns (DLOP, SLOP, SC, ARC) will be identified by the CAT with 90% accuracy. Data Filtering Results : The data was collected at 48 H z and immediately stored onto the o nboard memory. As such the data did have a significant amount of noise in addition to the constant resulting from gravity. Using Figure 23 : Unfiltered Frequency Spectrum ARC


74 MATLAB a Fast Fourier Transform was performed on the data to determine which frequencies in the data contained useful information. Looking at the frequency spectrum shown in Figure 23 , several things are immediately apparent. First there is a large spike at 0 Hz in both the Acceleration X and the Acceleration Y plots. This spike corresponds to a DC offset created by the acceleration due to gravity. Secondly , o nly Gyro Y has peaks that go above 5 Hertz. Then across the rest of the frequencies there appears to be just digital noise containing no relevant information. Interestingly , the FFT plots for the m agnetometer contains very few frequencies with any intensity and thus very little information. Frequencies can be seen at 0 Hz for these and distortions created by the wheelchair. Figure 24 : Unfiltered DLOP FFT


75 In general, the frequency graph shown in Figure 24 does correspond with the expected frequencies of normal human activity. The peak at one hertz likely describes the push rate of about one cycle per second. The peak frequently occurring at approximately two hertz likely corresponds to the rate of switching back and forth between propulsion and recovery phase. T h ese graphs also demonstrate that the sampling rate of the watch is well above the Nyquist sampling frequency. Interestingly each pu sh pattern appears to have different amplitudes for frequencies that occur in multiple patterns . Once again in Figure 25 the frequency spectrum for th e magnetic field sensor shows very few frequencies with minimal amplitudes . The gyroscope measuring angular rotation around the y axis again has a large frequency hump spanning between 3 and 7 hertz. The z axis is missing the peak at 0 hertz primarily due to the fact it remained perpendicular to the axis of gravity. This implies that at least during the testing, the participant performed very little pronation or supination of the wrist. Figure 25 : Unfiltered SC FFT


76 Figure 26 shows the frequency spectrum of the training data set for the SLOP pattern. Interestingly large peaks can be seen at 0 hertz for the acceleration in all three axes. Implying a greater amount of wrist rotation occurred . This increased wrist rotation caused each axis to experience the acceleration due to gravity. All Fast Fourier transform plots were taken before max min normalization was performed on any of the data sections. Figure 26 : Unfiltered SLOP FFT


77 A Butterworth bandpass filter was created using the M ATLAB filterbuilder command. The filter was designed to allow a bandpass for frequencies between approximately 0.4 Hz to 6 Hz. Removing the peak due to gravity and any higher frequencies that are a result of extraneous noise. The blue line in Figure 27 represents the frequency response of the band pass filter. No gain or decay is applied to the data for frequencies between approximately 1 and 6 Hz. On either side of this band, the gain appl ied to the system quickly removes these frequencies. The red line represents the phase shift of the data. T he phase shift is of little consequence for this filter. Figure 27 : Bandpass Frequency a nd Phase Shift Plot


78 Classifier Training Results: Random Forest A decision tree can also be used to make classifications based o n provided data. The provided training data was broken in to zones relating to set values . A decision tree makes decisions on new data by and then checks its list of if statemen ts and then returns the label assigned to the resulting leaf. In the example above , Figure 28 , the algorithm looks at variable x9 in the data provided . If variable x is less than 0.166 than the algorithm, then goes to the left branch of the tree. However, if x9 is greater than or equal to 0.166 then the algorithm then proceeds down the right branch of the tree. This is continued until a leaf is arrived a t . The class membership of the specific leaf is then returned as the classification of the data points. This method works best with features extracted from time series data. Due to the high probability of a single tree over fitting the data, a large numbe r of trees or a forest is often used. Each tree is built with different parameters and will likely classify some data points Figure 28 : Example Decision Tree


79 differently. The majority vote is chosen as the predicted class ; resulting in a higher probability of correctly classifying the dat a. MATLAB has built in functions for both creating a single decision tree as well as creating a forest. The command fitctree(data, label) generates a single decision tree. Then the command predict(tree, data) can be used to classify new data with the generated decision tree or forest . If a forest of decision trees is desired, the command TreeBagger(numTrees, data , label) can be used . Matlab will then generate the desired number of trees using the data and labels provided. Random Forest Without PCA: Just taking the raw non PCA transformed data a random forest was built using 2500 individual trees. Ninety five of each push pattern was used to train the algorithm and t wenty five of each pattern was withheld to be used as testing data to de t ermine the accuracy of the trained Random Forest. Figure 29 : Number of Trees and Out of Bag Error without PCA


80 Out of Bag error is the calculated probability that an observation in the training data will be misclassified by the random forest. When plotted against an increasing number of trees it is possible to determine the minimum number of trees necessary to achieve a specific level of accuracy. Figure 29 shows o nce the number of classification trees has reached approximately 100, the error ceases to drop and settles at approximately 5 %. Another measure of the performance of the Random Forest is using the withheld testing data with known classifications. This data is feed into the trained random forest and the percent correctly classified at each increase in the number of trees will be plotted. At some number of trees, the number correctly classified will stabilize. As shown in Figure 30 , this appears to occur at around 100 trees. Figure 30 : Percent Correct Vs Number of Trees


81 Zooming in on the first 200 trees as seen in Figure 31 the accuracy is greater than 95% once 20 trees have been used. The accuracy of the Random Forest ap pears to peak with the use of 80 trees at approximately 9 8 % . Even with just one decision tree the accuracy is approximately 76%. This shows that at least one or two classes are easily separable with the chosen variables. With less than 10 trees the accurac y already exceeded 90%. The variability in accuracy at each increase in number of trees occurs because of the random selection of data to train each tree. Figure 31 : Percent Correct Vs First 200 Trees


82 Additionally, the variables with the greatest importance to each classification can be determined. The letter combination after each feature in Figure 32 stands for, A for acceleration G for Gyroscope/angular rotation. Then the X, Y, Z are the corresponding directions for the acceleration or axis of angular rotation. SSC stands for slope sign change, RMS is the root mean squared, STD is the standard devi ation, MAD is the median absolute deviation, and the ZCR is the zero crossing rate. The variables that had the greatest impact on the decision trees was the average acceleration in the z direction, followed by the median value of the angular rotation aroun d the x axis, followed closely by the median acceleration in the Z direction. In general, the SSC, Kurtosis, Entropy, as well as the MAD variables appear to have the least impact on the ability to differentiate between the different classes. Figure 32 : Variable Importance for Classification Trees


83 A confusion matrix can be used to show how well each observation/push pattern was classified. As shown in Figure 33 , the SC, SLOP, and ARC patterns for the testing data were correctly identified every time. In one instance a single DLOP pattern it was incorrectly identified as the ARC pattern. Again, t here w as 25 of each propulsion cycle patterns with held during the training. This resulted in a total of 100 individual test classifications. Of these 100 patterns only 1 pattern was incorrectly identified. Effects of PCA on Random Forrest : Based on the perfor mance accuracy from random forest built without first transforming that data with PCA, 100 trees were repeatedly trained while increasing the number of principle components used to transform the training and testing data. Figure 33 : Random Forest Confusion Matrix (80 Trees)


84 As seen in Figure 34 , the most striking outcome of PCA was the immediate and drastic decrease in the accuracy of the Random Forest for the same number of trees. Even if all 54 principle components are considered, algorithm perform better. Figure 34 : Number of PC verses Accuracy


85 Strikingly, even when all 54 principle components are considered it took just over 500 trees before the accuracy appears to level off. The Out of Bag error levels off at 11% to 12%, this is significantly higher than the Out of Bag error not using P CA were the error levels off at 5%, with a significantly smaller number of decision trees. PCA does not appear to do anything for the Random Forest. Figure 36 : Out of Bag Error for all Principle Comp onents Figure 35 : Confusion Matrix (1700 Trees)


86 The trained random forest was then given the withheld testing data and Figure 35 shows the output classifications. The global accuracy is only 83% for this which is worse than the error seen in the Out of Bag error plot. The SC and SLOP patterns are classified correctly 92% of the time or 23 out of 25 times. DLOP suffers the worst from PCA with a 64% accuracy. The final test is to increase to the numbe r of trees until accuracy maxes out for different numbers of included principle components. Shown above in Figure 37 , increasing the number of principle components do es not significantly change accuracy above twenty five PC. Twenty five principle components result ed in the lowest classification accuracy at approximately 80% . The accuracy remains around 85% at 35, 45, and 54 principle components . Figure 37 : Percent Accuracy for PC at Increasing Number of Trees


87 Artificial Neural Network (ANN) MATLAB has a built in command to build and generate Artificial Neural Networks. The command patternnet(hiddenSize,trainFcn,performFcn) generates a neural network. The hiddenSizes input determines the number of hidden layers inside of the ANN . The trainFcn input determines what training function should be used to train the network. MATLAB has a variety of training methods . I f no option is provided , MATLAB defaults to the traingscg option which uses a scaled conjugate gradient method. The final option is the performFcn, which determines which performance function will be used to determine how well the network is performing . W hen left empty , MATLAB defaults to a cross entropy method. Cross entropy performance is calculated as the . Finally, the train command can be called with the generated neural network and given the training data and associated training labels. As with the Random Forest, the ANN was train ed on raw non PCA transformed data . Figure 38 : Effects of Hidden Layers on Percent Correct


88 An Artificial Neural Network was trained using the training data and then tested against the with held testing observations to find the percent accuracy . Each time an additional hidden layer was added to the neural network, the accuracy was re calculate d . After including approximately 10 hidden layers , as seen in Figure 38 , the ANN correctly classified the testing observations with a 95% accuracy. Based on the confusion matrix shown in Figure 39 , the neural net successfully classif ies the SC and the DLOP patterns with 100% accuracy with this training and testing set. While not perfect ly, the neural net performs well with the SLOP and ARC patterns. The neural net is more likely classify patterns as the SC pattern more than any other pattern, as it predicted 1 SLOP and 1 ARC pattern as a SC pattern. If different cycles were chosen as par t of the training or testing set some variability in the percent accuracy is probable. Figure 39 : ANN confusion Matrix (200 hidden layers)


89 When MATLAB trains a neural net it randomly with holds some of the training data to use as vali dation as well as for testing to determine if overfitting has occurred . The top left plot shows that the neural network can was able to successfully train the network to give the correct output for each of the training data samples . The closer the lines fa ll to the left and top axes the better the classification algorithm performs. By comparing the Training receiver operator curves (ROC) to the Test ROC, the Neural Net did not overfit to the training data. However, in calculating the sensitivity and specif icity for each individual pattern, MATLAB considers only the probability for th at specific class. If the probability is below a specific threshold it is short comings Figure 40 : Receiver Operator Curve for ANN


90 this threshold is affected by the classification probabilities of the other classes. For example, the target may be [0,0,1,0] and the output might be [.6, .1, .9,.01]. When the threshold drops below 0.6, technically two differ ent classes could be the desired class. Or in the case when the output is [.82,.2,.4,.1], the first class is the obvious guess to take. Yet the ROC plot would show it as a true positive if the threshold is lower than 0.4 . Support Vector Machine (SVM) Resul ts : A support vector machine works by finding a hyperplane that best separates all data points that belong to one class from all data points that belong to another class. The best hyperplane is defined as the plane with the greatest separation between eac h class such that there is an area/margin in which there are no data points for either class. Frequently data is not linearly separable, and a kernel must be used to move the data into a higher dimensionality in which there does exist a plane that can allo w for linear separation of the data. Once again MATLAB has built in functionality to build and use both binary and multiclass SVMs. If there are multiple classes, the command fitcecoc is used to build a multi class SVM. To perform multi classifications, m ultiple hyperplanes are generated. Each SVM attempting to separate one specific class from all other classes. The first test performed attempted to determine if PCA increas ed the accuracy of the SVM. One principle component was used to transform the data into a one axis problem, then two, three, until all possible principle components were considered.


91 Based on Figure 41 the accuracy maxes out after 27 principle components are considered and used to transform the data. After 27 principle components are used , the classification accuracy is 93% for any number o f additional principle components. This show s that each data class is poorly separated even by the principle components with the greatest amount of variability. No single variable accounts for a significant portion of the variance between patterns. Figure 41 : Number of Principle Component vs Percent Accuracy


92 When not using PCA, the support vector machine was able to classify the SC, SLOP, and ARC patterns with 96% accuracy. Only one pattern out of each 25 testing patterns were misclassified. In the same manner as the Neural Network, more patterns were guessed to be SC than any other pattern. DLOP had the greatest number of misclassifications with 4 of the 25 patterns being incorrectly identified . Classifications of DLOP as SLOP or SC make sense as the DLOP is partially a combination of both other patterns. When testing the SVM with d ifferent kernel functions only the linear method returns rational classification s are not capable of finding vectors that differentiate the data. Additionally, the GapTolerance, DeltaGradientTolerance, and Ka rush Kuhn Tucker Tolerance have little to no effect on accuracy when their values are not set to 0. Figure 42 : SVM Without Principle Components

PAGE 100

93 H ypothesis 2: Patients who use a greater number of arcing or single loop propulsion patterns as compared to subjects who use a greater number of semi cir cular or DLOP propulsion patterns will report significantly higher levels of upper limb joint pain. Specific Aim 2.1 : Using the activity tracker, compare propulsion pattern differences between a population of up to 20 MWU completing a simulated daily living environment. Controlled Testing Environment: The controlled course was set up according to the figure above. The participant started by first propelling over an ADA ramp with dimensions of 5.37 . The participant would go up the ramp and then immediately down an equally sized ramp. Then they had an distance to slow down and turn before propelling their chair over a thick carpet in length. Underneath the carpet , one layer of EVA foam w as added to increase cushioning and rolling resistance. The individual would turn again and going over a step ramp with dimension of ADA Ramp Carpet Short Ramp Tile Start Finish Figure 43 : Course Map

PAGE 101

94 . After going down from this ramp, a section of open tile was provi ded to allow time to slow and turn before traveling across an other flat tiled section in length . This circuit was repeated a total of three times . Data was collected from the wrist worn activity tracker as well as from the wheelchair IMU unit. Addition ally , a small camera was mounted to each manual wheelchair such that the arm with the custom tracker could be viewed during each propulsion pattern . This video was then reviewed by the author to visually determine which push patterns were employed during the course . Participants Demographics : Each participant was recruited and consented following an approved IRB protocol from the university of Colorado COMIRB 18 0115 . Only 3 individuals joined this study (2 female, 1 male). Table 5 : Basic Demographics Average (years) Standard Deviation (years) Age 53.67 18.15 Duration of Wheelchair Use 23.67 18.5 Figure 44 : Course Layout

PAGE 102

95 All three participants reported having experienced upper limb pain while using their manual wheelchairs while two were currently experiencing joint pain . All reported having experienc ed pain in their hands, wrists, and shoulders. Two also reported experiencing pain in their elbows. The participants were also asked to rank their pain from 0 to 10 with 0 representing no pain and 10 being so much pain they were unable to move . Reported pain was scored between 3 and 6 on the scale. Additionally, all three individua ls had spinal cord injuries somewhere between T3 and T9. In terms of normal level s of activity , each participant reported a different level of activity . Activity levels were reported as somewhat active, fairly active, and very active. Strikingly none of t he participants were aware of the various push pattern types . To avoid potentially affecting their natural push pattern , the participants were not provided any information on the various push patterns . During the health history questionnaire, the participa nts were only exposed the names of the push patterns. Wheelchairs: Each participant used their personal wheelchair during the testing . As such they each had different setup s , fitted specifically for them. One major similarity between each of the wheelcha irs was a lack of sides on the chairs. This lack of sides is noteworthy primarily because this prevented mounting the magnets to assist in determining push pattern timing. Additionally , most of the chairs did not have any large sections of Ferris materials . This problem was mostly due to the lack of sides on the wheelchairs. This prevented easy attachment of the wheelchair speed and angle detector . The wheelchair tracker was designed to mount to the sides of the wheelchairs using small magnets . To get aroun d this problem , the tracker was taped , with permission , to

PAGE 103

96 a spot on the chair within a short distance to the wheel. Unfortunately , the tracker stopped functioning during multiple tests and prevented data collection. Variations of wheel types was another immediately obvious difference between each wheelchair. One of the individuals had large tires designed for off roading. While the other two individuals had smaller tires. Additionally, one of the other individuals has a power assist built directly into th eir wheelchair . This individual was still allowed to enroll because they still had to push their wheelchair, just with less force. The specific recovery motions would likely remain unaffected by the power assist device. Pattern Classifications and Counts: Each participant wore the custom activity tracker on their right arm while they propelled their chair through the simulated environment. A video recording was set up to record the motions of their right arm during each propulsion and recovery stroke. Thi s video was then watched by the author and the start and stops of each push pattern were marked in the video . The same starts and stops were marked in the data along with the rater determined classification. To synchronize the data from the custom tracker with the video data, e ach participant was asked to hold the custom activity tracker still while the video was started. Once the video was started , the participants were told they could begin navigating the course. The mome nt the participant started to move was marked in the video. Similarly, the first change in acceleration was marked as the starting location in the data. Due to sampling rate differences between the video and IMU data, a conver sion between the video frame l ocation and the IMU data locations was necessary. This conversion was carried out by counting the number of video frames between the start and stop of each pattern . This value was then divided by the video frame rate and multiplied by the data sampling rat e. Then feature

PAGE 104

97 extraction was performed on each of these propulsion patterns and fed into the trained random forest and the trained neural net algorithms . These results are shown below. Participant 1 Results : Table 6 : Participant 1 Predicted and Actual Pattern Counts Pattern Actual Count Predicted Counts (Random Forest (80 Trees)) Predicted Counts (Neural Net (25 Layers)) Semi Circular (SC) 0 37 0 Single Loop Over (SLOP) 0 6 52 Double Loop Over (DLOP) 0 19 5 Arcing (ARC) 62 0 5 Total 62 62 62 Based on the video, the author determined that participant one only used the ARC pattern to propel the wheelchair. The ARC pattern was used regardless the surface type or ramp being traversed. While the same global classification was common throughout the simulated environment. The video revealed on such irregularity. During the simulated environment, the subject made m ultiple stutter An example of this stutter pattern is shown below in Figure 45 .

PAGE 105

98 During the normal arcing pattern, the hand grasps the wheel and in one motion moves from the grabbed location to the release point and then back along the handrim to the starting location. In transitioning to breaking from ARC pattern, the individual stops during the return stroke and does not complete the recovery phase , creating a stutter motion . The individual would release the push rim and then partially return up along the push rim before stopping to press against the rim to slow the wheelchair down . T his creates an additional two changes in direction of acceleration for this specific pattern observation , acting as an independent pattern . When treated as a single pattern, it greatly lengthens the total duration of the push pattern. The push pattern now contained multiple recovery and Due to the additional changes in accelerations, t his stutter pattern could almost be treated as two separate propulsion cycles . However, f or this study, the entire motion from the initial grab until the individual returned to the original starting position was considered one propulsion pattern. This avoided the issue of creating new classifications to describe each portion of the stuttered pattern. Regular ARC Breaking ARC Figure 45

PAGE 106

99 Little additional motions and stops like this were common throughout the entire course, especially when approaching sharp turn s . This motion was used to both slow the entire chair as wel l slow one wheel to turn the chair at corners. This stutter pattern was also used to finish propelling over a ramp before braking on the downward slope. Participant 2 Results : Table 7 : Participant 2 Predicted and Actual Pattern Counts Pattern Actual Count Predicted Counts (Random Forest (80 Trees)) Predicted Counts (Neural Net (25 Layers)) Semi Circular (SC) 1 3 0 Single Loop Over (SLOP) 0 6 32 Double Loop Over (DLOP) 17 54 29 Arcing (ARC) 47 2 4 Total 65 65 65 Start Release Pause Figure 46 : Propulsion to Breaking Transition

PAGE 107

100 Again , neither the Random Forest nor the Neural Net were able to correctly classify the given patterns. Most of the push patterns (72.3%) used were of the arcing type and (26.2%) were DLOP based on the authors judgement of the video. However, the R andom F orest p redicted that (83.1%) of the patterns were the DLOP pattern and only 3.1% were ARCing. The Neural Net predicted that 49.2% were SLOP and 44.6% were DLOP with 6.2% of the ARC variety. Participant two switched back and forth between two distinct patterns w ith their right hand. During tasks that involved starting, corners, or ramps, the participant used a clear Arcing pattern. However, once the participant got their wheelchair moving, they switched between a DLOP pattern and what appeared to be a mix between the SLOP and the DLOP pattern. This pattern was characterized by a flick of the hand up and away from the wheel at the end of propulsion phase. Then during the recovery phase, the individual proceed ed to move their hand i n a path level with the top of the push rim of the chair. Due to the initial release, the pattern appeared to take the form of a SLOP pattern . Then i nstead of lifting high above the push rim follow ed a straight path back to a natural grabbing location , in a manner more akin to the DLOP pattern. The basic hand trajectory is shown below in Figure 47 . patterns. Many of these irregular patterns were associated with a shift from propulsion to breaking or Figure 47 : Suppressed SLOP

PAGE 108

101 turning. Again, t hese instances were difficult to classify well into one of the four main propulsion patterns. For example, the individual would pause during recovery and perform a partial push before return ing to the normal starting contact location . For simplicity in classifying these irre gular patterns , a pattern was not considered completed until the hand returned to the starting wheel contact location. While maneuvering over a ramp , participant two significantly shortened their re covery phase compared to normal propulsions. This shorter recovery phase during ramp ascension assisted against gravity directly oppos ing the motion of the chair. To overcome gravity an individual can chang e propulsion cycle rate, increase force exerted , change location of initial contact, or use any combination. Participant two also exhibited different recovery patterns between their left and right hands during the testing. This difference was not noticed during casual observations before or after the te sting. This discrepancy likely occurred because the watch band would contact the wheel during the recovery phase. It is possible t he sensation subconsciously bothered the individual. To avoid contact of the watch band and push rim, the y adapted their recov ery pattern by allowing their hand to swing out wider from the chair. Furthermore , the act of wearing the watchband may have been enough to make the individual focus more on what they were doing with that hand. This increased focus may have caused them to over think their normal propulsion method. Figure 48 : Mixed SLOP/DLOP

PAGE 109

102 Participant 3 Results : Table 8 :Participant three Predicted and Actual Pattern Counts Pattern Actual Count Predicted Counts (Random Forest (80 Trees)) Predicted Counts (Neural Net (25 Layers)) Semi Circular (SC) 0 8 0 Single Loop Over (SLOP) 0 82 3 7 Double Loop Over (DLOP) 0 0 49 Arcing (ARC) 90 0 4 Total 90 90 90 Participant three was the oldest and by far the weakest of the individuals who volunteered for the study. This participant only used the arcing pattern and very clearly attempted to use as little range of motion as possible . Some of the smallest motions occurred when addit ional force was required due to starting motion , crossing the carpet, or ramp interference. The smaller motions and weaker propulsion can clearly be seen in the increased number of pushes required to go through the course 3 times. Participant three took 90 pushes to complete the course where participants one and two completed it in 62 and 65 pushes respectively. The types of misclassifications above demonstrate an importance of finding additional features that relate to the hands direction of travel during recovery phase. Classifications based on the video, are unlikely to confuse ARC and SLOP patterns. A rater has the advantage of using the wheel as a reference

PAGE 110

103 point for the hand position. Features that describe the direction or change of directions may in crease the robustness of the overall algorithm. General Simulated Environment Results: Based on the video data collected during th e simulated environment, the types of motions required to propel a wheelchair are far more complex than the described normal four push patterns. Every participant made continuous adjustments to starting grip location s and releasing points . A s the participant prepared for or navigated over obstacle s , there were changes to their push pattern s . Most studies describing the four main push patterns examine data from steady state propulsion. The simulated environment did not allow for long periods of consistent steady state pushing, which created constant needs for adjustments. A minimum of at least three additional classes of propulsion methods should be added to better describe the collected data. The first class would be a breaking motion. Frequently individuals would move through the propulsion phase and then stop mid recovery or even mid propuls ion and press their hand into the push rim to slow the chair down. As mentioned above, t his added additional challenges in determining when one full push cycle was completed. The next two classes that should be studied are turning left and turning right. T his is important because this is another motion that disrupts the normal propulsion patterns causing the individuals to vary their push patterns. However, during turning maneuvers, an MWU will perform different tasks with each hand. One hand pushes while t he other completely stops or slows the rotation of the other wheel.

PAGE 111

104 Specific Aim 2 .2 Results: Using the activity tracker compare, at minimum, full day use (10 hours) of kinematic differences between the before mentioned group in their normal routines. By examining Tables 3, 4, and 5 , the current algorithm is not capable of accurately predicting other individual s pattern s. The rater classifications and the algorithm predicted classifications agreed less than 10% of the time. As such, it is not possible to determine the success or accuracy of the second hypothesis (Patients who use a greater number of arcing or single loop propulsion patterns as compared to subjects who use a greater number of semi circular or DLOP propulsion patterns will report higher l evels of upper limb joint pain). Data was collected from individuals in their normal daily life. However, this data has not been feed through the classifier due to the poor performance of the classifier. Any results that would be returned would have such a large margin of error that the r esults could not be trusted.

PAGE 112

105 CHAPTER VI D ISCUSSION The results from the initial training data collection are encouraging. All three algorithms employed were able to perform with greater than 90% accuracy. Any of the three methods could potentially be used as a classifier. However, the Random Forest as well as the ANN performed better at classifying all four types of patterns as compared to the SVM. The Random Forest need ed approximately 80 trees before its accuracy maxes out, while the neural network required approximately 20 layers for optimal classifications. Due to the small number of trees or hidden layers necessary, there is a strong likelihood of being able to deploy the algorithms on to a wrist worn activity tracker. In a worst c ase scenario, the algorithms could be easily deployed on to a mobile phone app. Data from the tracker could be continuously . After syncing the classifications and push pattern counts could be determined. Once the inco ming data was correctly structured, the predict function for the Artificial Neural Network with 25 hidden layers could make a classification prediction on average every 0.008 seconds. The Random Forrest at 80 trees could make a classification prediction on average every 0.005 seconds. These results will have a significant amount of variance based on other processes being run on the computer as well as the type of processor being used. Performing feature extraction took approximately 0.00126 seconds per prop ulsion cycle. These results were obtained from MATLAB running on an Intel i5 clocked at 3.2 Ghz. This is not the first study to have noticed other extraneous push patterns being used during normal propulsion. This is also not the first study to have also n oticed the difficulty of classifying certain push patterns. non

PAGE 113

106 in creating a wearable tracker. Several important questions must be considered including, what does a comp leted push pattern with irregularities look like ? How should a stuttered recovery pattern be counted? Is the stuttered pattern one or two distinct patterns? What is the correct classification when an individual pushes and then wipes their hand on their pan t leg before finishing the recovery phase (as seen in the video) ? There truly is an almost limitless number of possibilities for the recovery phase. Any sort of true real time classifier will require a method of classifying an irregular pattern as Full characterization of manual wheelchair propulsion would require collecting wheel contact locations, forces, and times . This data could be used to help separate sections of the IMU data corresponding to the recovery and propulsion phase s . Eac h of the four push patterns has a similar propulsion phase, with some possible differences in average initial contact and release locations. There exists the possibility that joint pain/injury occurs most in individuals who have some type of variation in t hese contact angles and times. The large amount of error seen in classifying data from other individuals is a function of environmental differences between the training and testing areas . The training data was collected while the manual wheelchair was prop elled up and down a long mostly flat, tiled hallway. As such a rhythm could be established and followed for each push pattern . The subject specifically concentrat ed on performing each specific pattern. This additional concentration may have influenced the consistency of each push. In contrast, t he simulated environment forced m ultiple changes of speeds, directions, and resistances . These changes interfer ed with the development of a rhythm. Individuals going through the simulated environment were likely using their natural propulsion techniques . The individual who provided the algorithm training data was at times using unfamiliar patterns that could feel unnatural.

PAGE 114

107 During collection of the initial train ing data, it was clear that a significant amount of focus was being used to ensure conformance to the requested propulsion pattern. This additional focus will affect how naturally all motions are performed. This could cause exaggeration of some aspects of the different patterns and other subtle changes . These exaggerations could affect the ability to classify patterns from individuals who are naturally following a pattern they have been using for years. Another possible cause of the poor performance could b e from differences in strength , size, and weight of the participants. For example, a strong individual using the same pattern as a weak er individual of the same size would be able to accelerate their chair at a much greater rate. This higher acceleration w ould affect many of the features extracted such as by increasing the mean and median. The classifications were performed on data comprising of both the propulsion and recovery phases . This could have made the algorithms more likely to be affected by differ ences in wheelchair setup that impacted contact angles and locations . Rate of travel was not measured or attempted to be held constant between participants. One method of increasing velocity is spending less time in recovery by moving their arms at a highe r rate back to the starting contact point. A higher velocity over the same distance, necessitates larger accelerations which would create individual unique IMU characterization data. The average accelerations in the X and Y directions over the entire course of testing was calculated for each participant. The X and Y d irections are shown below only because they most directly correspond to the motions observed in the 4 main push patterns as compared to angular rotation rates . The results are shown below in Table 9 . Even when the raw data was normalized using the max/min normalization, a similar splat tering of results could be seen.

PAGE 115

108 Table 9 : Average Accelerations Over Entire Test Course Average Acceleration X Average Acceleration Y Participant 1 3.011 5.25 Participant 2 5.529 5.948 Participant 3 4.766 6.376 Participant 1 and Participant 3 only used ARCing patterns during the simulated environment. However, the average accelerations for both individuals during this time are not similar. Participant 1 was significantly younger than participant 3 ; yet , somehow appears to have had lower average accelerations even with requiring less pushes than participant 3 to complete the course. All three individuals experienced accelerations the peaked above 18 , and all three had reported experiencing joint pai n. When examining raw accelerometer data, large peaks could be clearly seen which directly relate d to the rapid change between propulsion and recovery. These peaks existed in every single one of the propulsion patterns . C onsistently , switching between pro pulsion and recovery resulted in the greatest acceleration. When observing position graphs of the normal four push patterns, some idea of expected acceleration changes can be found. With additional studies it may be possible to relate these changes of dire ctional travel to the frequencies seen in the FFT plots.

PAGE 116

109 In Figure 49 above the orange dots mark locations where the direction of travel during recovery changes. To have multiple changes in direction of travel , several acceleration changes must occur. When exam ining the FFT plots, each plot had a spike at 1 Hz. This likely occurs due to each push occurring approximately 1 second apart. Then the other peaks are likely function s of how many changes occur in the direction of travel of their hand . This likely causes the peaks that are often seen at 2, 3 and 4 Hz . Another factor that will influence various frequencies in the acceleration plots is the amount of effort needed to drive the wheels. T forces beg in to decelerate the wheels. At the point of contact, a greater amount of force will be required to begin accelerating the wheel . As the wheel rotational rate increases, the forces acting on the wheel eventually balance out. During this time the accelerati will likely adjust accordingly. The process of distilling individual push patterns from the raw data is incredibly labor intensive and has a high likelihood of error in finding the true start and stop of each pattern. Additionally, th ere was some unknown and unquantifiable time offset between the true video and data sync time. This very likely contributed to the error seen in classifying the propulsion patterns from the simulated Figure 49 : Changes in Accelerations

PAGE 117

110 environment study. For example, even a small temporal of fset could lose the large peak accelerations from the propulsion cycle. The loss of these peaks could result in significantly different extracted feature values, which would decrease classification performance. Future Work/Next Steps: A statistically sign ificant number of individuals should be enrolled in another study where all individuals provide training data under ideal conditions and perform in the simulated environment. Then the IMU data should be used to individually train and test the classifiers, resulting in an accuracy per individual. Additionally, IMU data from the other users should be tested against each individually trained algorithm. Then each algorithm should be trained on data from all enrolled individuals and then tested against withheld data with known classifications and known provider . This would create a much larger database of the push patterns and could allow for a more robust algorithm. Time and effort should be invested in ensuring that the data and video were correctly temporally synchronized . Correct timing would allow for accurate determination of the starts and stops of each push pattern. Finally, data collected from the simulated environments could be tested against the trained algorithm as well as used as part of t he training data sets. Another future step is to move the entire classification algorithm to the wrist worn activity tracker and attempt real time classifications. To best accomplish this goal, a robust method of determining the start and stops of each pattern should be created. This would allow the activity tracker to only attempt classifications once the entire pattern has been completed and avoid classifying non propulsion related motions. Depending on the method chosen, it may be possible to classify the recovery phase distinct from the propulsion phase. Other common techniques used during manual wheelchair propulsion should be included in the possible classification system. These include but are not limited to breaking, turning left, and turning

PAGE 118

111 right. Beyond tracking the propulsion patterns, themselves, it would be beneficial to track uphill and downhill velocity, grade, and duration. Doing so should contribute valuable information pertaining to the effects of daily activities performed by manual wheelchair users. Combining time synchronized data from a system such as the Smart Wheel system and this CAT would result in another useful study. This would connect contact times and forces to the accelerations and arm. The combined data set would provide a much fuller picture of the variables that play a role in the development of upper limb joint pain. It is possible that fully identifying the causes of upper limb joint pain will require this increased level of monitoring. One additional study would be to classify various wheelc hair setups using the same algorithms. MWU could be recruited to propel a customized chair with adjustable features. They c ould be asked to either just propel their chairs normally , or to perform a specified pattern at each of the wheelchair adjustments. T hen the CAT would be able to both predict wheelchair setup as well as the specific propulsion patterns used. This level of data analysis could be of great use to clinicians in prescribing wheelchairs. Limitations: As this was a pilot/feasibility study, no t enough individuals were enrolled for either the training data group, or the simulated environment group to draw complete conclusions. No data was collected from an individual who had never experienced joint pain . Furthermore, each of the individuals were in their own chairs, each customized specifically for them . This could have affect the natural contact locations and impact ed features of the recovery phase. w as not able to also perform in the simulated environment. Having been able to do so would have given additional insight into the results of the simulated environment. It would have provided helpful results

PAGE 119

112 to have tested the trained algorithm against the d ata acquired during the simulated course. Doing so would have helped demonstrate how much the course really affected the classifications as opposed to variations between each own pattern. It is entirely likely, that the algorithm would still have been unable to accurately classify these patterns, just due to the differences between constrained/controlled environments and more natural propulsion. Additionally, this study only collected and trained the algorithm on the four main push patterns under ideal conditions . This training did not consider other common ly used movement s during manual wheelchair propulsion . These movements , as discussed above , include actions such as breaking and turning. Due to the method of video and IMU data capture, difficulties arose in synchronizing the data with the video. Some of the error found in classifying the data from the simulated environment is likely due to this problem. Removing this error should increase classification accuracies and greatly simplify so me of the data analysis process. A greater amount of confidence could be given to each algorithm accuracy, knowing that the algorithm attempted a classification on the complete IMU data set from one propulsion.

PAGE 120

113 CHAPTER VII C ONCLUSIONS It is possible to take IMU data , specifically Accelerometer and Angular Rate of Rotation , from a wrist worn activity tracker and classify the propulsion patterns used with greater than 90% accuracy . Unfortunately, an algorithm trained on data collected by one individual does not appear to be capable of accurately classifying patterns of other MWU naturally propelling their own chairs. As such no claims can be made about differences in types of patterns used by individuals with and without joint pain. There is potential fo r a future custom activity tracker to aid clinicians in preventing upper limb joint injury and pain by monitoring upper limb kinematics using IMU.

PAGE 121

114 R EFERENCES : [1] [Accessed: 13 Jun 2017]. [2] K. A. Curtis, G. A. Drysdale, R. D. Lanza, M. Kolber, R. S. Vitolo Arch. Phys. Med. Rehabil. , vol. 80, no. 4, pp. 453 457, Apr. 1999. [3] Michael L. Boninger, MD; Jeffrey D. Towers, MD; Rory A. Cooper, PhD; Brad E. Dicianno; and Michael C. Semantic imaging abnormalities in individuals with Boninger Towers/f5ba67e2f98936d8d1eed4ac84212651f059bd9a. [Accessed: 20 Oct 2017 ]. [4] Clin. Biomech. , vol. 30, no. 9, pp. 927 932, Nov. 2015. [5] . [6] [7] E. National Academies of Sciences, The Promise of Assistive Technology to Enhance Activity and Work Participation . 20 17. [8] N. H. White and N. [9] Common Conditions Requiring Rehabilitation in the United States: Stroke, Spinal Cord Injury, Traumatic Brain Injury, Multiple Sclerosis, Osteoarthritis, Rhe umatoid Arthritis, Limb Loss, and Arch. Phys. Med. Rehabil. , vol. 95, no. 5, pp. 986 995.e1, May 2014. [10] of a seated wheelchair or mobility sc ooter for people with traumatic brain injury or spinal cord Aust. Occup. Ther. J. , vol. 60, no. 6, pp. 378 386. [11] Aust. Occup. Ther. J. , vol. 50, no. 1, pp. 30 39 . [12] Posture and performance in C5 Spinal Cord , vol. 38, no. 7, pp. 425 434, Jul. 2000. [13] tion of J. Rehabil. Res. Dev. , vol. 35, no. 2, p. 210, 1998. [14] J. L. Mercer, M. Boninger, A. Koontz, D. Ren, T. Dyson Clin. Biomech. , vol. 21, no. 8, pp. 781 789, Oct. 2006.

PAGE 122

115 [15] Arch. Phys. Med. Rehabil. , vol. 83, no. 5, pp . 718 723, May 2002. [16] Propulsion Patterns on Natural Surfaces During Start Arch. Phys. Med. Rehabil. , vol. 90, no. 11, pp. 1916 1923, Nov. 2009. [17] Arch. Phys. Med. Rehabil. , vol. 83, no. 6, pp. 776 783, Jun. 2002. [18] I. Vujaklija et al. Are the Performance Front. Neurorobotics , vol. 11, 2017. [19] g a triaxial [20] Hum. Mov. Sci. , vol. 18, no. 5, pp. 589 602, Oct. 1999. [21] tic variability among 12 motion analysis laboratories [Online]. Available:!/content/playContent/1 s2.0 S0966636208003573? 6636208003573%3Fshowa ll%3Dtrue&referrer=. [Accessed: 12 Jun 2018]. [22] K. S. Lee and M. Int. J. Occup. Saf. Ergon. , vol. 0, no. ja, pp. 1 49, Jun. 2018. [23] one linear projection of low J. Physiol. , vol. 587, no. Pt 24, pp. 5925 5938, Dec. 2009. [24] K. C. McGil J. Neurosci. Methods , vol. 149, no. 2, pp. 121 133, Dec. 2005. [25] Marieb, Elaine and Hoehn, Katja, Human Anatomy & Physiology, 9th Edition , 9th ed. Pearson. [26] [Accessed: 28 Feb 2018]. [27] Outcomes o Arch. Phys. Med. Rehabil. , vol. 89, no. 2, pp. 260 268, Feb. 2008.

PAGE 123

116 [28] K. T. Asato, R. A. Cooper, R. N. Robertson, and J. F. IEEE Trans. Biomed. Eng. , vol. 40, no. 12, pp. 1320 1324, Dec. 1993. [29] ResearchGate . [Online]. Available: _Ambulation. [Accessed: 13 Jun 2017]. [30] s: Research and Med. Eng. Phys. , vol. 28, no. 9, pp. 905 915, Nov. 2006. [31] Disabil. Rehabil. Assist. Technol. , vol. 11, no. 3, pp. 177 187, Apr. 2016. [32] [33] IEEE Trans. Rehabil. Eng. , vol. 4, no. 3, pp. 182 187, Sep. 1996. [34] Over Cross Assist. Technol. , vol. 23, no. 1, pp. 42 51, Mar. 2011. [35] Appl. Ergon. , vol. 30, no. 6, pp. 555 563, Dec. 1999. [36] Int. J. Ind. Ergon. , vol. 44, no. 6, pp. 851 856, Nov. 2014. [37] P. Gorce Clin. Biomech. , vol. 27, no. 1, pp. 7 15, Jan. 2012. [38] al Wheelchair J. Phys. Ther. Sci. , vol. 26, no. 2, pp. 313 318, Feb. 2014. [39] C. Y. Tsai, C. J. Lin, Y. C. Huang, P. C. Lin, and F. wheel camber on the kinematics of upper extremity Biomed. Eng. OnLine , vol. 11, p. 87, Nov. 2012. [40] Y. S. Yang, A. M. Koontz, S. J. Yeh, and J. Arch. Phys. Med. R ehabil. , vol. 93, no. 4, pp. 654 659, Apr. 2012. [41] Ann. Rehabil. Med. , vol. 35, no. 6, pp. 897 906, Dec. 2011.

PAGE 124

117 [42] M. Giner Pascual, M. Alcanyis Alberola, L. Millan González, M. Aguilar Rodríguez, and F. Querol, Int. J. Rehabil. Res. , vol. 34, no. 4, pp. 282 289, Dec. 2011. [43] S. Wei, S. L. Huang, C. J. Jiang, and J. C. Chi Clin. Biomech. , vol. 18, no. 6, pp. S46 S52, Jul. 2003. [44] magicwheels . [Online]. Available: m. [Accessed: 28 Feb 2018]. [45] Speed Geared Manual Wheelchair Propulsion on Arch. Phys. Med. Rehabil. , vol. 88, no. 12, pp. 1622 1627, Dec. 2007. [46] Omid Jahanian, Alyssa J. Schn Shoulder Joint Kinematics and Stroke Cycle characterstics during Geared And Standard Manual eelchair_seating/jahanian.html. [Accessed: 24 Jan 2018]. [47] wheelchair wheel to reduce propulsive muscular demand during ramp ascent: Analysis of muscle Clin. Biomech. , vol. 25, no. 1, pp. 21 28, Jan. 2010. [48] [49] els | THE RIGHT WAY TO ROLL TM [Accessed: 28 Feb 2018]. [50] Arch. Phys. Med. Rehabil. , vol. 79, no. 1, pp. 36 40, Jan. 1998. [51] Throughout Wheelchair Propulsion in Able Bodied and Persons With Paraplegia: An Integrative J . Biomech. Eng. , vol. 131, no. 2, pp. 021015 021015 8, Dec. 2008. [52] Spi nal Cord , vol. 42, no. 2, pp. 91 98, 2004. [53] A. Faupin, P. Gorce, P. Campillo, A. Thevenon, and O. Rémy Clin. Biomech. , vol. 21, no. 6, pp. 560 566, Jul. 2006. [54] efficiency of two commercial lever J. Rehabil. Res. Dev. , vol. 50, no. 10, pp. 1363 1372, 2013.

PAGE 125

118 [55] activated power Arch. Phys. Med. Rehabil. , vol. 86, no. 3, pp. 380 386, Mar. 2005. [56] Am. J. Phys. Med. Rehabil. , vol. 89, no. 3, pp. 225 234, Mar. 2010. [57] review on the pros and cons of using a pushrim activated power Clin. Rehabil. , vol. 27, no. 4, pp. 299 313, 2013. [58] T. A And Stroke Frequency Differences Between Manual Wheelchair Propulsion And Pushrim Activated Power J. Spinal Cord Med. , vol. 26, n o. 2, pp. 135 140, Jan. 2003. [59] Disabil. Rehabil. Assist. Technol. , vol. 7, no. 6, pp. 459 463, Nov. 2012. [60] Spinal Cord , vol. 36, no. 6, 1998. [61] M. L. Bon Arch. Phys. Med. Rehabil. , vol. 78, no. 4, pp. 364 372, Apr. 1997. [62] J. L. Davis, E. S. G dimensional kinematics J. Rehabil. Res. Dev. , vol. 35, no. 1, pp. 61 72, Jan. 1998. [63] M. L. Boninger, R. A. Cooper, M. Arch. Phys. Med. Rehabil. , vol. 80, no. 8, pp. 910 915, Aug. 1999. [64] Iowas State University Environmental Health and Safety . [Online]. Available: [Accessed: 29 Jan 2018]. [65] J. Nagy et al. Wi Top. Spinal Cord Inj. Rehabil. , vol. 18, no. 2, pp. 140 142, 2012. [66] J. Spinal Cord Med. , vol. 28, no. 5, pp. 434 470, 2005. [67] P. Frost et al. der tendinitis in relation to shoulder loads in monotonous repetitive Am. J. Ind. Med. , vol. 41, no. 1, pp. 11 18, Jan. 2002.

PAGE 126

119 [68] M. M. B. Morrow, M. G. Van Straaten, N. S. Murthy, J. P. Braman, E. Zanella, and K. D. Zhao, BioMed Res. Int. , vol. 2014, 2014. [69] Based Quantitative Ultrasound in Manual Wheelchair Users, Am. J. Phys. Med. Rehabil. Assoc. Acad. Physiatr. , vol. 89, no. 5, pp. 390 400, May 2010. [70] Neurocomputing , vol. 167, no. Supplement C, pp. 52 60, Nov. 2015. [71] Grained IEEE Internet Comput. , vol. 19, no. 5, pp. 26 35, Sep. 2015. [72] S. Majumder, T Sensors , vol. 17, no. 1, Jan. 2017. [73] Comput. Secur. , vol. 39, pp. 137 144, Nov. 2013. [74] S. Mu Gait Posture , vol. 18, no. 1, pp. 114 125, Aug. 2003. [75] I. P. Machado, A. discovery from triaxial accelerometer sensor: Non supervised learning sensitivity to feature Inf. Process. Manag. , vol. 51, no. 2, pp. 204 214, Mar. 2015. [76] IEEE Trans. Biomed. Eng. , vol. 56, no. 3, pp. 871 879, Mar. 2009. [77] Neurocomputing , vol. 120, no. Supplement C, pp. 633 644, Nov. 2013. [78] R J. Biomech. , vol. 41, no. 11, pp. 2438 2445, Aug. 2008. [79] D. M. Karantonis et al. time hum an movement classifier using a triaxial IEEE Trans Inf. Technol. Biomed. , pp. 156 167, 2006. [80] based human activity re Pattern Recognit. Lett. , vol. 73, Jan. 2016. [81] Wbracelet Exclusive Jewelry , 31 Jul 2017. .