Machine learning-based multi-domain model for SOT composite score prediction using quiet standing COP data

The composite score (CS) obtained via the sensory organization test (SOT) serves as a key indicator of overall balance control. However, traditional assessment methods require expensive instrumentation, ample physical space, and multiple assessment conditions, thereby restricting their accessibility. To overcome these limitations, we developed a machine learning-based multi-domain model that predicts CS directly from quiet standing center of pressure (COP) data. A cohort of 302 participants completed SOT assessments, and their quiet standing COP signals were used for both model training and evaluation. Our model incorporates both time-demographic and frequency-domain features into a comprehensive feature set, providing a more holistic representation of postural control when compared with single-domain models. The model demonstrated strong predictive performance with R² = 0.990, RMSE = 0.632, and MAE = 0.338, offering a significant improvement over single-domain models (p < 0.001). These results indicate that applying machine learning to quiet standing COP data can yield highly accurate CS estimates, presenting an efficient and accessible method for balance assessment with potential applications in fall-risk evaluation, rehabilitation, and monitoring of disease progression.