A data-driven scoring model for automated assessment of balance rehabilitation exercises (Preprint)

BACKGROUND Balance rehabilitation programs constitute the most common treatment for balance disorders. Nonetheless, lack of resources and of highly expert physiotherapists barriers for patients to have individualized rehabilitation sessions and the required in person monitoring of the rehabilitation program. Therefore, balance rehabilitation programs are often transferred to the home environment, with a high risk of the patient mis-performing the exercises or failing to follow the programme at all. OBJECTIVE Holobalance is a persuasive coaching system with the capacity to offer full-scale rehabilitation services at home. Holobalance comprises of several modules, from rehabilitation program management to Augmented Reality (AR) coach presentation. One of the modules is the scoring module, which assesses the performance of the patient using wearable devices. METHODS The data-driven scoring module described in the present work is based on an extensive dataset (~1300 rehabilitation exercise sessions) collected during the Holobalance pilot study. Its scope is to be used as a training and testing dataset for training machine learning (ML) models, which can infer the scoring components of all physical rehabilitation exercises. RESULTS The results of the trained model improved the performance of the scoring module, in terms of more accurate assessment of a performed exercise, when compared to a rule-based scoring model deployed at an early phase of the system (k-statistic value of 15.9% for the sitting exercises, 20.8% for the standing exercises and 26.8% for the walking exercises). CONCLUSIONS Finally, the resulting performance of the model resembles the threshold of the interobserver variability, enabling the trustworthy usage of the scoring module to the closed-loop chain of the Holobalance coaching system.