PancreaNomogram: A multimodal fusion nomogram incorporating ultrasound radiomics, deep embeddings, and clinical variables for diagnostic assessment of type 2 diabetes

This study aimed to develop and validate an interpretable, multimodal predictive model integrating pancreatic ultrasound radiomics, deep-learning embeddings, and structured clinical features to estimate individualized risk of insulin-induced hypoglycemia in adults with type 2 diabetes. In this retrospective multicenter cohort study, 4786 insulin-treated adults with type 2 diabetes mellitus were screened across six tertiary centers, resulting in 3160 eligible participants after predefined exclusions. Standardized pancreatic ultrasound examinations were acquired and preprocessed using intensity normalization, CLAHE enhancement, and wavelet-based speckle suppression. Radiomic features (n = 1268) and deep-learning embeddings extracted from five convolutional and transformer-based architectures were filtered using reproducibility thresholds (ICC ≥0.75), followed by univariate screening, correlation pruning, and LASSO regularization. Clinical variables were harmonized using a unified data dictionary. Multimodal integration was performed via late feature-level fusion. Data were stratified and split into training, validation, and independent test cohorts (70/15/15) to enable robust multicenter evaluation. Model performance was assessed using AUC, calibration metrics, Brier score, and Decision Curve Analysis, and an interpretable logistic-regression nomogram was constructed. On the independent test cohort, the multimodal feature-level fusion model achieved the highest predictive performance for insulin-induced hypoglycemia (AUC = 0.953, 95% CI 0.941–0.965), outperforming radiomic-only (0.882), deep-only (0.918), and clinical-only (0.812) models. Calibration demonstrated excellent agreement between predicted and observed risks (Brier score = 0.071; calibration slope = 0.98). Decision Curve Analysis showed consistently higher net benefit across clinically relevant thresholds (0.20–0.60), corresponding to the prevention of approximately nine unnecessary interventions per 100 patients compared with clinical-only strategies. The final Radiomics–Deep–Clinical nomogram preserved strong discrimination (AUC = 0.924) while providing transparent and clinically interpretable individualized risk estimates. This study presents a robust, and clinically interpretable multimodal framework that substantially improves individualized prediction of insulin-induced hypoglycemia and supports precision insulin management.