Development and validation of a multimodal interpretable machine learning model for the identification of osteoporosis in patients with type 2 diabetes mellitus: a multicenter retrospective study.

Background: Osteoporosis is a prevalent yet underdiagnosed complication in patients with type 2 diabetes mellitus (T2DM), significantly increasing the risk of fractures and mortality. However, current screening tools are limited by low accuracy and lack of generalizability. This multicenter cross-sectional study aimed to develop and validate a multimodal interpretable machine learning (ML) model integrating multimodal data to identify and classify osteoporosis risk in T2DM patients, with a focus on clinical translatability. Methods: We retrospectively enrolled 1,002 T2DM patients from two tertiary hospitals in China. After exclusions, Cohort 1 (n = 852) was used for model development and internal validation, and Cohort 2 (n = 126) served as an independent external validation set. Multimodal data included demographics, laboratory tests, abdominal CT-derived parameters (e.g., skeletal muscle index at L3, SMI-L3), and composite metabolic indices (e.g., lymphocyte-to-HDL ratio [LHR], Metabolic Score for Visceral Fat [METS-VF]). Core predictors were selected using univariate analysis, Least Absolute Shrinkage and Selection Operator (LASSO) regression, and the Boruta algorithm. Seven ML algorithms were compared, with model performance evaluated by area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, F1-score, and area under the precision-recall curve (AUPRC). The SHapley Additive exPlanations (SHAP) method was applied for model interpretability. Results: Seven predictors were identified: age, hemoglobin, neutrophil count, uric acid, LHR, SMI-L3, and METS-VF. The eXtreme Gradient Boosting (XGBoost) model demonstrated superior performance, achieving an AUC of 0.877 (95% confidence interval [CI]: 0.830-0.923) in the training set and 0.911 (95% CI: 0.879-0.943) in the external validation set. SHAP analysis revealed age, hemoglobin, and uric acid as the top contributors. Calibration and decision curve analyses confirmed the model's clinical utility. Conclusion: This study presents a robust, multimodal interpretable ML model for osteoporosis risk prediction in T2DM using routinely available clinical data. By integrating multimodal features and providing transparent predictions via SHAP, the model offers a proof-of-concept tool for opportunistic screening to facilitate targeted dual-energy X-ray absorptiometry (DXA) referral, although prospective clinical validation is needed before its clinical utility can be established.