A novel inertial measurement unit-based device for capturing upper extremity biomechanics

Exercise adherence can be poor for patients in physical therapy (PT). Unsupervised exercise may contribute to poor adherence as patients may have low exercise self-efficacy, feel unmotivated, or simply do not perform exercises correctly. This dissertation proposes a cheap wearable device that utilizes inertial measurement units (IMUs) to capture kinematics during upper extremity (UE) exercises. The first aim of this study was to identify key ergonomic and graphic features that would likely improve device adoption and assess participant receptiveness towards the device. The second aim was to validate the device against marker-based motion capture using traditional validation techniques and trend symmetry analysis (TSA). The third aim was to evaluate machine learning models for classifying nine different upper extremity exercises, based upon kinematic data captured from an IMU-based device. 50 participants were recruited to perform nine common UE exercises with the device. Joint angles were calculated for the shoulder, elbow, and wrist. Participants were given an orientation to the device and shown conceptual software graphics. Participants responded positively to the device and graphics but found the device bulky and cumbersome. Participants were able to interpret kinematic analysis despite having limited biomechanics knowledge. Device range of motion (ROM) measures were most accurate for sagittal and frontal plane measures (ROM measure differences ranged from 3.3°-18.6°) with worse performance on transverse plane movements (ROM measure differences ranged from 6.0°-60.8°). TSA scores were high (mean = 0.956) even when ROM agreement was low. TSA shows promise as a validation technique, but additional research is needed to establish cut points to define “good” TSA results. UE exercises could be classified with accuracy as high as 98.6% using random forest models using flattened kinematic data. Down sampling had minimal impact on classifier performance (precision = 0.952, 5 Hz). Accuracy remained greater than 90% when training splits were decreased to 10%. Classifier performance decreased (96% accuracy with 90% training split) when using stratified train-test splits, and overtraining may pose a problem with small samples. The findings of this research can provide guidance on the design of future wearables to augment PT and development of a revised device and companion smartphone application for clinical testing