Explainable artificial intelligence on safe balance and its major determinants in stroke patients

Area of Science:

• Neuroscience
• Artificial Intelligence
• Rehabilitation Medicine

Background:

• Stroke significantly impairs mobility and balance, necessitating accurate prediction of recovery.
• Predicting safe balance is crucial for patient rehabilitation and independence post-stroke.
• Current prediction methods may not fully leverage multimodal patient data.

Purpose of the Study:

• To develop and validate an explainable artificial intelligence (AI) model for predicting safe balance in stroke survivors.
• To identify key predictors of safe balance, including clinical, neurophysiological, and diffusion tensor imaging (DTI) data.
• To compare the performance of AI models against traditional logistic regression.

Main Methods:

• Retrospective analysis of data from 92 first-time stroke patients.
• Utilized random forest models with Shapley Additive Explanation (SHAP) for prediction and interpretability.
• Included Berg Balance Scale (BBS) scores, Fugl-Meyer Assessment (FMA) scores, muscle strength, and DTI properties (fractional anisotropy) as predictors.

Main Results:

• Random forest models achieved high accuracy (AUC 91-92%), outperforming logistic regression (AUC 87-92%).
• Key predictors for BBS at 3 months included BBS at 1 month, FMA, and ipsilesional corticospinal tract fractional anisotropy.
• Key predictors for BBS at 6 months included FMA, BBS at 1 month, and lower limb muscle strength.

Conclusions:

• Safe balance recovery after stroke is strongly associated with initial motor function and corticospinal tract integrity.
• Explainable AI, integrating DTI data, enhances the prediction of safe balance post-stroke.
• The developed AI model provides insights into factors influencing balance recovery, aiding personalized rehabilitation.