Machine Learning for Personalized Physiotherapy: Predicting Patient Recovery and Treatment Efficacy

Abstract

Personalized physiotherapy depends on objectively knowing what a patient is doing during exercise, yet clinicians still rely largely on intermittent observation and self-report. This study examines whether wearable inertial sensor data can support that sensing layer using machine learning. We used a public inertial measurement unit (IMU) dataset of 200 sessions across eight standardized physiotherapy exercises, from which fifteen statistical features (mean, standard deviation, minimum, maximum, median of accelerometer, gyroscope, and magnetometer magnitude) were engineered and used to train tuned Random Forest and XGBoost classifiers. XGBoost achieved 55.0% accuracy and a macro-F1 of 0.54 across the eight classes (range: 0.22-0.89), with magnetometer-derived features consistently the strongest predictors. t-SNE and UMAP projections corroborated the supervised confusion patterns, showing partial but incomplete class separability. These results show that low-cost wearable sensors combined with interpretable ensembles can meaningfully, if imperfectly, discriminate physiotherapy exercises without laboratory motion capture. We frame this exercise-recognition capability as a necessary precursor to, not a substitute for, models that directly predict patient recovery and treatment efficacy, and outline the research gap and roadmap for connecting objectively measured movement quality to validated clinical outcomes.