A Bi-LSTM based approach for Compensation detection during robotic stroke rehabilitation therapy

Abstract

Stroke is a common cause of brain damage and death due to blocked blood flow to the brain. Prolonged strokes increase the risk of irreversible brain damage and bodily harm. Stroke rehabilitation entails objective setting, robust exercise, comprehensive care, and purpose-driven training, occasionally including robotic systems for increased therapy. However, unsupervised patients often compensate with unaffected joints and muscles, hindering rehabilitation outcomes. This study detects these compensatory movements using LSTM and Bi-LSTM models. The dataset includes 3-D joint position trajectories of the upper body over time. Results show the Bi-LSTM model, with accuracies of 97% for healthy subjects and 98% for stroke patients, outperforms other models, maintaining consistency in both 4-class and 2-class problems.