Automated Movement Assessment in Stroke Rehabilitation

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Artificial Intelligence in Healthcare

Background:

• Stroke survivors often require long-term rehabilitation, with home-based programs being crucial for recovery.
• Current home rehabilitation assessment methods can be limited by cost, accessibility, and objective measurement.
• Automated assessment systems are needed to provide consistent and reliable feedback for upper extremity stroke rehabilitation.

Purpose of the Study:

• To develop and validate a cyber-human methodology for the Semi-Automated Rehabilitation At the Home (SARAH) system.
• To create a hierarchical model for automated segmentation of stroke survivor movements and performance scoring.
• To leverage low-cost, unobtrusive video-based sensing for effective home-based rehabilitation assessment.

Main Methods:

• A hierarchical model combining expert knowledge (task structure) with data-driven techniques (HMM, Decision Tree, Transformer, MS-TCN).
• Utilizing RGB images and raw kinematics, processed through transformer and Multi-Stage Temporal Convolutional Network (MS-TCN) architectures.
• Developing a sequence combining complementary algorithms to encode movement hierarchy information for robust analysis.

Main Results:

• Achieved 85% accuracy in per-frame movement labeling.
• Reached 99% accuracy in classifying movement segments.
• Demonstrated 93% accuracy in assessing task completion during rehabilitation.

Conclusions:

• The SARAH system provides accurate and robust automated assessment for upper extremity stroke rehabilitation at home.
• The proposed cyber-human methodology effectively handles noisy and variable data from low-cost video capture.
• This approach has potential for broader application in other rehabilitation contexts, such as lower extremity training.