Real-Time Activity Recognition After Stroke; Machine Learning Development For A Wearable Exercise-Cueing System
Date of Award
Summer 8-17-2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Biomedical Engineering
First Advisor
Robert Scheidt
Second Advisor
Kimberly Bassindale
Third Advisor
Maura Casadio
Fourth Advisor
Whitney Morelli
Abstract
Stroke commonly results in persistent upper-extremity motor impairment, limiting independence in activities of daily living and reducing quality of life. Although high-repetition, task-specific practice is important for motor recovery, supervised rehabilitation time is limited, leaving substantial downtime that may be underused for therapeutic activity. Souvenir, a bilateral wrist-worn activity monitoring and cueing system, was previously developed to cue progressive upper-extremity activities during unsupervised downtime after stroke. Although cue delivery has shown to increase activity, Souvenir has yet to interpret cue responses in a way that provides activity feedback to the user as a method to increase motivation. This thesis developed and evaluated supervised machine learning models for classifying upper-extremity cue responses based on kinematic data collected during unsupervised use of Souvenir. Bilateral upper extremity kinematic data were collected at 12.5 samples per second using low-cost, commercial, inertial measurement unit (IMU) systems coupled to a smartphone. IMU data were segmented around cue events, processed into movement features, labeled according to observed cue-response activity, and used to train selected candidate machine learning classifiers. Initial model evaluations compared candidate model architectures using accuracy, precision, recall, F1-score, and confusion matrices. A finer-resolution hyperparameter analysis was then performed on top-performing models using accuracy and estimated computational cost per inference to evaluate feasibility for mobile deployment. Among the candidate models, random forest and multilayer perceptron models achieved comparable peak accuracy during initial evaluation (88% and 86% respectively). A Random Forest architecture was ultimately selected as the preferred model for mobile deployment because Pareto optimization found that a Random Forest configuration with just 3 trees and a maximum tree depth of 9 decisions could achieve an accuracy of 86.88% with a maximum of only 27 operations (OPs) per inference. These results demonstrate that supervised machine learning can classify task-specific upper-extremity cue responses from bilateral wrist-worn IMU data with high fidelity at low computational cost. Future integration of the selected model into Souvenir may support real-time activity recognition and provide real-time task-specific feedback to motivate poststroke users to perform specific, cued, therapeutic, exercise activities.