Improving Accelerometry-Based Measurement of Functional Use of the Upper Extremity After Stroke: Machine Learning...

Area of Science:

• Biomedical Engineering
• Rehabilitation Technology
• Movement Science

Background:

• Wrist-worn accelerometry offers objective monitoring of upper-extremity functional use.
• However, accelerometry can detect nonfunctional movements, creating ambiguity in monitoring results.

Purpose of the Study:

• To compare machine learning algorithms against standard methods (counts ratio) for improved accuracy in detecting functional activity.
• To assess the efficacy of machine learning in distinguishing functional from nonfunctional movements using accelerometry data.

Main Methods:

• Synchronous accelerometry and video recording of healthy controls and individuals with stroke performing unstructured tasks.
• Human annotation of video frames for ground truth functional/nonfunctional activity.
• Development and comparison of machine learning algorithms and the counts ratio method.

Main Results:

• The counts ratio method showed poor correlation with ground truth (r = 0.48) and high average error (52.7%) due to nonfunctional movements.
• The best-performing intrasubject machine learning model achieved 92.6% accuracy and high correlation (r = 0.99) with ground truth in stroke patients.
• The best intersubject model demonstrated 74.2% accuracy and a correlation of r = 0.81 with ground truth.

Conclusions:

• Standard counts ratio method is inaccurate for reflecting functional upper-extremity activity in stroke patients.
• Machine learning algorithms significantly enhance the accuracy of functional activity detection from accelerometry data.
• Future development should focus on creating clinical tools based on machine learning for objective patient monitoring.