Enhancing stroke recovery assessment: A machine learning approach to real-world hand function analysis.

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Machine Learning in Healthcare

Background:

• Stroke survivors often experience hand weakness, impacting daily function and quality of life.
• Traditional accelerometer metrics for upper limb (UL) use lack clinical discrimination.
• Existing methods struggle to capture meaningful differences in post-stroke recovery.

Purpose of the Study:

• To develop a machine learning (ML) method for categorizing post-stroke upper limb performance.
• To align accelerometer data analysis with clinically validated Action Research Arm Test (ARAT) scores.
• To improve objective assessment of real-world UL function.

Main Methods:

• Utilized continuous 24-hour triaxial accelerometer data for UL movement analysis.
• Applied a deep neural network to extract features directly from raw accelerometer data.
• Categorized participants into five performance groups based on ML-learned features and ARAT scores.

Main Results:

• Achieved 97% classification accuracy in categorizing UL performance aligned with ARAT scores.
• The ML-based groupings were non-overlapping and clinically meaningful.
• Significantly outperformed traditional demographic and heuristic feature models (66% accuracy).

Conclusions:

• A novel ML framework accurately classifies clinically relevant UL function from accelerometer data.
• This method provides a more precise and objective assessment of post-stroke hand use.
• Potential applications include personalized rehabilitation planning and outcome monitoring.