Concurrent validity of machine learning-classified functional upper extremity use from accelerometry in chronic...

Area of Science:

• Rehabilitation Medicine
• Machine Learning in Healthcare
• Neuroscience

Background:

• Quantifying real-world upper extremity (UE) use is crucial for stroke recovery assessment.
• Traditional methods like video analysis are labor-intensive, and self-report measures can be subjective.
• Machine learning (ML) offers a potential for objective, scalable measurement of UE activity.

Purpose of the Study:

• To validate ML-derived estimates of real-world UE use against video-based ground truth in stroke survivors.
• To assess the correlation of ML-derived UE use with established measures of impairment, function, and dexterity.
• To determine the feasibility of ML for measuring UE recovery in clinical trials.

Main Methods:

• Participants (n=31) with chronic stroke wore accelerometers on both arms.
• Video recording captured participants performing daily activities to establish ground-truth UE use.
• A random forest classifier was trained on accelerometry data to estimate UE use ratio.

Main Results:

• ML-estimated UE use ratio strongly correlated with video-derived ratios (Bland-Altman plots showed excellent agreement).
• ML-derived use also correlated significantly with the Action Research Arm Test (ARAT), Fugl-Meyer (UEFM), and Nine-Hole Peg Test (9HPT).
• ML-derived UE use explained 83% of the variance in UE motor performance, capturing similar constructs to clinical tests.

Conclusions:

• The developed ML approach provides a valid and accurate measure of functional UE use in stroke survivors.
• Its accuracy, validity, and minimal footprint make it suitable for large-scale UE recovery monitoring in rehabilitation.
• This technology can enhance objective outcome measurement in stroke clinical trials.