Machine Learning-Based Estimation of Upper Extremity Function in Stroke Rehabilitation Using Body-Worn Inertial...

Area of Science:

• Rehabilitation Medicine
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Stroke is a primary cause of disability, notably affecting upper extremity (UE) function.
• Current clinical assessments like the Action Research Arm Test (ARAT) are vital but resource-intensive.
• There's a need for accessible, efficient methods to track patient recovery and tailor interventions.

Purpose of the Study:

• To develop and validate a machine learning model for estimating ARAT scores using wearable inertial sensors.
• To reduce the time and expertise required for assessing UE function in stroke survivors.
• To enable more frequent and objective monitoring of rehabilitation progress.

Main Methods:

• A machine learning model was trained using data from ActiGraph wearable sensors placed on the wrists and waist of 23 chronic stroke patients.
• The model predicted total ARAT scores based on a minimal set of UE tasks, selecting one item from each sub-test (grasp, grip, pinch, gross movement).
• Nested cross-validation was employed for optimizing item selection, feature selection, and model hyperparameters.

Main Results:

• The optimized machine learning model accurately estimated total ARAT scores with a median absolute error of 3.81 points.
• The model achieved a high coefficient of determination (0.93), indicating strong predictive performance.
• SHapley Additive explanations (SHAP) identified key sensor-derived features contributing to the ARAT score predictions.

Conclusions:

• Wearable sensor data combined with machine learning offers a promising approach for objective and efficient UE function assessment in stroke rehabilitation.
• This technology can facilitate more frequent progress monitoring across the continuum of care.
• The developed model has the potential to decrease the workload for clinicians and improve accessibility of standardized assessments.