Advancing Telerehabilitation with AI: Predicting Balance Scores from Early Sessions

Telerehabilitation systems have transformed the management of balance and gait impairments, particularly in older adults with chronic low back pain. This study integrates multimodal balance exercises, cognitive retraining, and augmented reality within the HOLOBALANCE system. Using data from an eight-week rehabilitation program, predictive models based on machine learning (XGBoost and AdaBoost) were developed to estimate final balance outcomes using as few as 3 to 6 sessions. The results indicate that XGBoost performs optimally for shorter prediction horizons (i.e., 3-4 sessions) with RMSE values of 0.292 and 0.351, and R² values of 0.014 and 0.216, respectively, whereas AdaBoost achieves superior accuracy at longer horizons (i.e., 6 sessions) with RMSE of 0.21 and a lower MAPE of 13.5%. Key predictive features, including training progression, physical performance, and pain levels, were consistently validated across models. This predictive capability allows clinicians to tailor interventions early, optimizing therapy outcomes and resource allocation. These findings highlight the potential of predictive modeling to enhance rehabilitation strategies by enabling early intervention and personalized adjustments tailored to individual patient needs. The use of AI algorithms bridges the gap between advanced technological interventions and traditional rehabilitation, providing a powerful tool for improving clinical decision-making and its outcomes in postural control rehabilitation.