Accelerometry and the Capacity-Performance Gap: Case Series Report in Upper-Extremity Motor Impairment Assessment...

Area of Science:

• Neuroscience
• Rehabilitation Medicine
• Biomedical Engineering

Background:

• Chronic stroke often leads to impaired functional hand use.
• Monitoring real-world hand use is crucial for effective rehabilitation.
• Existing methods for assessing hand function can be limited in ecological validity.

Purpose of the Study:

• To investigate the accuracy of machine learning (ML) and convolutional neural network (CNN) models in predicting real-world functional hand use.
• To assess the feasibility of using lab-trained models for monitoring hand function in chronic stroke patients.
• To identify patterns of learned non-use and their correlation with impairment severity.

Main Methods:

• A case series involving 4 participants with chronic stroke.
• Collection of wrist-worn accelerometry data in both laboratory and at-home settings.
• Concurrent video recording and frame-by-frame annotation using the FAABOS scale as ground truth.
• Training and evaluation of Random Forest ML and CNN models on in-lab data to predict at-home hand use.

Main Results:

• A capacity-performance gap was observed, with participants using their impaired hand less at home than in the lab.
• Random Forest ML models achieved high accuracy (0.80-0.90) in classifying at-home hand use for mild and severe impairments.
• Moderate impairments showed lower classification accuracy (0.62) for at-home hand use.
• CNN models demonstrated accuracy comparable to Random Forest classifiers.

Conclusions:

• Lab-trained ML models show feasibility for monitoring real-world hand use in stroke survivors.
• These models can help identify emerging patterns of learned non-use, enabling timely interventions.
• This approach holds promise for enhancing outpatient stroke rehabilitation and promoting functional recovery.