Objective:

To develop and validate a multimodal interpretable machine learning model for identifying and classifying osteoporosis risk in patients with type 2 diabetes mellitus (T2DM).

Approach:

• Study Design: A multicenter cross-sectional study involving 1,002 T2DM patients from two tertiary hospitals in China.
• Data Collection: Multimodal data included demographics, laboratory tests, abdominal CT-derived parameters, and composite metabolic indices.
• Model Development: Cohort 1 (n = 852) was used for model development and internal validation, while Cohort 2 (n = 126) served as an independent external validation set.
• Predictor Selection: Core predictors were selected using univariate analysis, LASSO regression, and the Boruta algorithm.
• Model Evaluation: Seven ML algorithms were compared, with performance evaluated by AUC, accuracy, sensitivity, specificity, F1-score, and AUPRC.
• Interpretability: The SHAP method was applied for model interpretability.

Key Findings:

• Seven predictors identified: age, hemoglobin, neutrophil count, uric acid, lymphocyte-to-HDL ratio, skeletal muscle index at L3, and metabolic score for visceral fat.
• The eXtreme Gradient Boosting (XGBoost) model achieved an AUC of 0.877 (95% 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 indicated age, hemoglobin, and uric acid as the top contributors to the model.

Interpretation:

The model demonstrates robust performance for osteoporosis risk prediction in T2DM patients using routinely available clinical data.

Limitations:

• The study is cross-sectional, limiting causal inference.
• Prospective clinical validation is needed to establish the model's clinical utility.

Conclusion:

The study presents a multimodal interpretable ML model for osteoporosis risk prediction in T2DM, offering a proof-of-concept tool for opportunistic screening.