Interpretable Source-Aware Bone-Muscle Fusion Framework for Multicenter Prediction of Osteoporotic Vertebral...

Osteoporotic vertebral fracture (OVF) is underrecognized, and bone mineral density (BMD) alone may be insufficient for prediction. We developed the multisource fusion model for osteoporotic vertebral fracture (MSF-OVF), an interpretable source-aware framework integrating clinical-demographic variables, BMD, and lumbar MRI-derived ventral and dorsal paraspinal muscle (PM) representations. This retrospective multicenter study included 2680 participants aged > 40 years from five centers (one training, one internal validation, four external test cohorts). Deep features from four PM regions were extracted using vision foundation model-guided transfer learning and integrated with age, sex, body mass index, and BMD through source-aware latent fusion. Performance was evaluated using the area under the receiver operating characteristic curve (AUC), net reclassification improvement (NRI), calibration, and interpretability analyses. MSF-OVF achieved an AUC of 0.881 (95% CI, 0.854-0.909) in internal validation and 0.851-0.917 across four external cohorts. Compared with the clinical-demographic and BMD baseline, MSF-OVF improved AUC by 0.046-0.093 and significantly outperformed it in internal validation and most external cohorts (DeLong P ≤ 0.034). Continuous NRI was 0.268-1.034 across all validation cohorts (P ≤ 0.011). MSF-OVF also outperformed single-muscle models (AUC, 0.454-0.736; P ≤ 0.009). After training-derived logistic recalibration, the internal validation Brier score was 0.140, with a calibration-in-the-large (CITL) of - 0.168 and calibration slope of 0.926. Class activation maps mainly highlighted PM regions. Automated segmentation took several seconds and inference 57.71 ± 18.66 ms per participant. Source-aware fusion of clinical-demographic, BMD, and PM MRI information may support computer-aided multicenter OVF prediction.