Advancing Telerehabilitation with AI: Predicting Balance Scores from Early Sessions.

Area of Science:

• Rehabilitation Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Telerehabilitation systems are crucial for managing balance and gait impairments in older adults with chronic low back pain.
• Multimodal approaches integrating physical exercises, cognitive training, and augmented reality show promise.

Purpose of the Study:

• To develop and validate machine learning models for predicting final balance outcomes in telerehabilitation.
• To assess the efficacy of early prediction using limited rehabilitation sessions.

Main Methods:

• Utilized data from an eight-week telerehabilitation program involving the HOLOBALANCE system.
• Developed and compared XGBoost and AdaBoost machine learning models for outcome prediction.
• Identified key predictive features such as training progression, physical performance, and pain levels.

Main Results:

• XGBoost models excelled at shorter prediction horizons (3-4 sessions) with specific RMSE and R² values.
• AdaBoost demonstrated superior accuracy at longer horizons (6 sessions) with lower RMSE and MAPE.
• Key features consistently validated across models, highlighting their predictive importance.

Conclusions:

• Predictive modeling enables early intervention and personalized adjustments in rehabilitation.
• Machine learning enhances clinical decision-making and optimizes resource allocation in postural control rehabilitation.
• AI integration bridges technological advancements with traditional rehabilitation practices.