Predicting online motor learning after stroke in lower limb task using machine learning.

Area of Science:

• Neurorehabilitation
• Machine Learning in Medicine
• Stroke Recovery

Background:

• Online motor learning is vital for effective rehabilitation and functional recovery post-stroke.
• Predictors of online motor learning after stroke are not well understood.
• Traditional statistical methods have limitations in identifying complex predictors.

Purpose of the Study:

• To leverage Machine Learning (ML) to identify key predictors of online motor learning in stroke survivors.
• To compare the predictive performance of ML models (XGBoost) against traditional logistic regression.
• To establish a framework for understanding individual differences in motor learning potential during rehabilitation sessions.

Main Methods:

• One hundred and seven stroke survivors underwent comprehensive assessments (sociodemographic, stroke characteristics, health, functional, cognitive, physical).
• Online motor learning was measured using a goal-directed ankle task.
• XGBoost and logistic regression models were applied to predict online motor learning capacity.

Main Results:

• The XGBoost model significantly outperformed logistic regression, demonstrating high precision (0.82), recall (0.78), F1 score (0.80), and AUC (0.84).
• Key predictors identified by ML include logical memory, DGT-backward, selective attention, functional capacity index, and physical capacity.
• Online motor learning post-stroke is influenced by a combination of cognitive, functional, and physical factors.

Conclusions:

• Online motor learning after stroke is a multidomain process, significantly influenced by cognitive and functional capacities.
• Machine learning effectively identifies complex, multidomain predictors of motor learning in stroke survivors.
• These findings support the development of precision rehabilitation strategies tailored to individual learning potential.