Feasibility of accelerometer-based prediction of postural balance in schoolchildren using machine learning models

Objective assessment of balance during childhood is essential for supporting motor development, preventing falls, and identifying potential impairments. Traditional clinical tests, such as the Flamingo and balance beam, are widely used but remain limited by subjectivity and lack of precision. Accelerometry offers quantitative measures, and its integration with machine learning models provides an opportunity to explore predictive assessment in pediatric populations. This study evaluated the feasibility of predicting clinical balance test outcomes in schoolchildren using accelerometric data from both static and dynamic tasks. A cross-sectional study was conducted with 90 children aged 6 to 12 years. Accelerometric signals were recorded in three axes and magnitude during static tasks (eyes closed, eyes open, and unstable surface) and during gait. Outcomes included the number of supports in the Flamingo test and the distance covered in the balance beam test. Several machine learning models, including linear regression, penalized regression, k-nearest neighbors, support vector regression, and random forest, were applied using 5-fold cross-validation. Models showed modest but consistent predictive accuracy for the Flamingo test, particularly with static tasks, with random forest, support vector regression, and k-nearest neighbors performing best. Prediction of balance beam outcomes was poor across all models. These findings suggest that accelerometry-based machine learning is feasible for predicting balance performance in children, especially for the Flamingo test, supporting its potential as a digital tool for screening and educational health applications.