Comparison of Machine Learning approaches for Classifying Upper Extremity Tasks in Individuals Post-Stroke.

Area of Science:

• Neurorehabilitation
• Biomedical Engineering
• Machine Learning

Background:

• Stroke frequently causes upper extremity (UE) motor dysfunction, impacting daily activities.
• Tracking UE movement recovery is crucial for assessing intervention effectiveness.
• Person-specific variability in post-stroke UE movements complicates automated gesture recognition.

Purpose of the Study:

• To identify an optimal set of sensor-extracted features for classifying UE movements in individuals post-stroke.
• To evaluate the efficacy of machine learning models for recognizing unimanual and bimanual gestures in an atypical population.
• To compare user-dependent and user-independent classification models.

Main Methods:

• Utilized sensor data from 20 individuals post-stroke and 20 age-matched controls.
• Extracted and selected sensor-based features for gesture classification.
• Employed a random forest classifier to categorize unimanual and bimanual gestures during task performance.

Main Results:

• An optimal set of fewer than 100 features achieved high classification performance.
• The random forest classifier demonstrated effectiveness across both post-stroke and control groups.
• Both user-dependent and user-independent models yielded robust results.

Conclusions:

• Machine learning, particularly with a random forest classifier and a reduced feature set, can effectively classify upper extremity movements in individuals post-stroke.
• This approach shows promise for objective assessment of UE motor function and recovery monitoring.
• The findings support the use of sensor-based machine learning for personalized rehabilitation strategies.