Data-Driven Quantitation of Movement Abnormality after Stroke.

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Artificial Intelligence in Medicine

Background:

• Stroke frequently impairs upper extremity (UE) motor function, posing significant challenges for clinical assessment.
• Current methods for measuring UE movement abnormalities lack precision and practicality, hindering effective therapeutic tracking and treatment.
• There is a critical need for objective, reliable, and user-friendly tools to quantify motor deficits post-stroke.

Purpose of the Study:

• To develop and validate a novel approach combining high-dimensional motion capture with out-of-distribution (OOD) detection for precise UE movement analysis.
• To assess the feasibility of using deep learning models trained on healthy data to identify abnormal movements in chronic stroke survivors.
• To correlate model performance with clinical measures of motor impairment.

Main Methods:

• Utilized wearable inertial measurement units (IMUs) to capture high-dimensional upper body motion data from healthy individuals and chronic stroke survivors.
• Developed and trained deep learning models exclusively on data from healthy subjects to classify functional movement primitives.
• Employed OOD detection principles to evaluate model confidence (prediction probabilities) when analyzing unseen data from both healthy and stroke groups.

Main Results:

• Models trained on healthy data exhibited high confidence for healthy motion but significantly lower confidence for stroke data, indicating OOD detection of abnormality.
• Decreased model confidence strongly correlated with the severity of motor impairment in stroke survivors.
• Motion data from the paretic upper extremity (UE) had a greater impact on model confidence than trunk movement data.

Conclusions:

• Out-of-distribution detection applied to high-dimensional motion capture data offers a precise and pragmatic method for identifying clinically meaningful UE movement abnormalities in chronic stroke.
• This approach has the potential to enhance objective assessment, therapeutic monitoring, and personalized rehabilitation strategies for stroke survivors.
• The findings highlight the utility of AI-driven OOD detection in bridging the gap between technological innovation and clinical needs in neurorehabilitation.