Multi-modal machine learning framework for automated osteoporosis diagnosis from knee X-rays: Emphasizing reproducibility, clinical utility via DCA, and nomogram construction

The objective of this study was to develop a comprehensive, automated diagnostic framework for osteoporosis using knee X-ray imaging by integrating radiomic features, deep learning-derived features, and clinical variables, and to validate its clinical applicability through a multi-center dataset. A retrospective, multi-center study was conducted with 3265 knee X-ray images collected from eight imaging centers. Radiomic features (n = 107) were extracted using PyRadiomics, and deep features were obtained from EfficientNetB0 and Vision Transformer models. Feature reproducibility was assessed using the Intraclass Correlation Coefficient (ICC), and features with ICC >0.75 were retained. LASSO regression was applied for feature selection, while fused features underwent dimensionality reduction using principal component analysis (PCA) before selection. Machine learning models including XGBoost, Random Forest, SVM, and Logistic Regression were trained and validated using cross-validation on training (80 %) and internal test (20 %) splits, with an independent external test cohort for final evaluation. Interpretability was assessed via attention maps. A final clinical nomogram was constructed by incorporating fused imaging scores with key clinical factors. The fused feature model achieved the best performance, with XGBoost yielding an AUC-ROC of 95.7 % in training, 95.0 % in internal validation, and 94.3 % in external testing. Vision Transformer features alone outperformed EfficientNetB0 and radiomics, but fusion consistently enhanced predictive ability. The nomogram demonstrated strong calibration and clinical utility in Decision Curve Analysis (DCA). Attention maps confirmed the alignment of model focus with radiologically relevant regions, supporting interpretability and clinical integration potential. The proposed multi-modal framework offers a robust, interpretable, and clinically applicable tool for osteoporosis screening from knee X-rays. Prospective validation and expansion into fracture risk prediction are warranted.