Predicting Response to Exercise Therapy in Adolescents With Spinal Curvature Abnormalities: A Randomized Controlled Trial Using Machine Learning (Preprint)

BACKGROUND Background: Adolescence is a critical period for spinal and neuromuscular development, during which abnormal spinal curvature may progress rapidly and lead to long-term musculoskeletal dysfunction. Exercise therapy is widely recommended as a non-surgical intervention; however, substantial individual variability in treatment response limits its clinical effectiveness. Although multidimensional data on body composition and spinal function are routinely collected in schools and rehabilitation clinics, these data are rarely integrated into intervention decision-making. Current screening and treatment selection still rely largely on visual assessment and simple angular measurements, and validated tools for identifying adolescents most likely to benefit from specific exercise therapies are lacking. OBJECTIVE Objective: This study aimed to evaluate the effects of a 12-week spiral muscle chain training (SPS) and combined exercise therapy incorporating proprioceptive neuromuscular facilitation (PNF), and to develop an interpretable machine learning–based predictive model to support personalized exercise therapy planning for adolescents with abnormal spinal curvature. METHODS Methods: The data for this study were derived from a 12-week randomized controlled trial of exercise therapy. A total of 125 middle and high school students with abnormal spinal curvature were recruited from schools and randomly assigned to a spiral muscle chain training group (n = 61) or a combined exercise therapy group (n = 64). All interventions were conducted offline. Baseline and post-intervention assessments of body composition and spinal health were performed using standardized clinical measurements. Singular value decomposition–based principal component analysis (SVD-PCA) was applied to extract principal components representing spinal mobility and balance. These components, together with demographic and clinical indicators, were used to construct predictive models using four machine learning algorithms: K-nearest neighbors (KNN), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost). Model performance was evaluated, and SHapley Additive exPlanations (SHAP) were used to interpret the optimal model. RESULTS Results: Both exercise therapies significantly improved spinal curvature, spinal mobility, and head, shoulder, and pelvic balance, with combined exercise therapy demonstrating superior efficacy. The reduction in angle of trunk inclination (ATI) was greater in the combined therapy group(P