Abstract WMP27: Augmenting Early Prediction Of Post-Rehabilitation Ambulation Ability And Risk Of Falling Status For Recovering Stroke Patients Using Wearable Sensors And Machine Learning

Abstract

Introduction: Inertial measurement units (IMU) can capture objective biomarkers of recovery during everyday activity paving the way to early prognosis and personalized therapy in post-stroke rehabilitation. In this study, we evaluated the performance of IMU data incorporated into machine learning models to predict discharge abilities compared to models based on traditional clinical variables. Methods: 32 subacute stroke patients admitted to inpatient rehabilitation performed clinical tests while wearing IMU sensors placed on the pelvis and both unaffected and affected-side shanks (US, AS). Variables collected were patient information (PI), clinician-scored functional assessments (FA), signal-processed IMU data during the 10-Meter Walk Test (IMU 10MWT ) and the Berg-Balance Scale (IMU BBS ). Patients were identified as household or community ambulators (Amb), and with high or normal risk of fall (RoF) based on discharge functional scores. A class-weighted L1-penalized logistic regression model was trained using nested leave-one-out cross-validation on admission data to predict discharge Amb and RoF statuses. The impact of sensor data on model performance was assessed on three feature sets: PI+FA, PI+IMU 10MWT/BBS , PI+FA+IMU 10MWT/BBS . Evaluation metrics included: weighted F1 score (WF1), accuracy, log-loss, and feature importance. Results: The PI+FA+IMU BBS , and both the PI+IMU 10MWT and PI+FA+IMU 10MWT models, outperformed the PI + FA benchmark model for RoF and Amb prediction, respectively, across all metrics (Figure 1). US gyroscopic and AS/US acceleration were relevant for Amb prediction, while the admission BBS score and pelvic acceleration were important to RoF prediction. Conclusions: Incorporation of simple IMU data from admission can improve discharge functional predictions over a model using PI and FA. Early, accurate predictions of post-stroke recovery could facilitate more personalized and efficient rehabilitation strategies.