Multimodal prediction of chronic ankle instability and exploratory treatment response using longitudinal delta-radiomics, wearable biomechanics, and conformal prediction

Background Chronic ankle instability (CAI) develops in about 40% of patients after a lateral ankle sprain, yet no validated tool separates rehabilitation responders from surgical candidates at initial presentation. Methods In a multicenter cohort of 1420 patients (development: n = 1040; external test: n = 380), we combined baseline and 6-week MRI (delta-radiomics), 7-day free-living IMU biomechanics, and patient-reported outcomes. A Feature Tokenization Transformer (FT-Transformer) was compared with LightGBM, XGBoost, logistic regression, and TabPFN v2 using nested 5-fold cross-validation. A federated model (Fed-GBDT; ε = 1.0 differential privacy) enabled privacy-preserving deployment, and a conformal prediction nomogram provided 90%-coverage uncertainty intervals. Results On external testing, the FT-Transformer achieved an AUC of 0.869 (95% CI: 0.843–0.895) and an accuracy of 80.0% for CAI diagnosis, and an AUC of 0.838 (95% CI: 0.789–0.887) for treatment-response stratification (exploratory). Delta-radiomics (ΔAUC +0.023, 95% CI: 0.013–0.033) and IMU features (ΔAUC +0.015, 95% CI: 0.005–0.025) added independent value. The federated model matched the centralized model (ΔAUC −0.003), and the imaging + IMU-only variant (AUC 0.843) exceeded two radiologists (AUC 0.803 and 0.779; different rating scales). Nomogram coverage was 91.4%. Conclusions This framework combines longitudinal delta-radiomics, wearable biomechanics, and federated learning with calibrated uncertainty for CAI prediction. Treatment-response stratification is hypothesis-generating and needs prospective multicenter validation before clinical use.