Combining MRI dual-sequence radiomics and clinical parameters predicts post-hypertriglyceridemic acute pancreatitis diabetes mellitus

Figures

• 

  Figure 1. Flowchart of the patient selection process.

• 

  Figure 2. Radiomics workflow for the development, evaluation, and interpretation of the predictive model. The analytical pipeline consists of five main stages: (1) Segmentation: Three-dimensional regions of interest (ROIs) were manually delineated on dual-sequence MR images (CE-T1WI and T2WI). (2) Feature Extraction: High-throughput radiomic features were extracted from original and transformed images (Log and Wavelet filters). (3) Feature Selection: Optimal predictive features were identified through a stepwise dimensionality reduction process comprising variance thresholding, SelectKBest, and LASSO regression. The optimal LASSO regularization parameter (λ) was determined via 10-fold cross-validation. (4) Model Analysis and Evaluation: The predictive performance of the combined model was comprehensively assessed using ROC curves, PR curves, and calibration curves with Brier scores, while clinical utility was evaluated via DCA. (5) Model Interpretation: The internal decision-making logic of the models was elucidated using the SHAP framework, encompassing global feature importance, summary beeswarm plots, and local individual waterfall plots. HTG-AP, hypertriglyceridemia-associated acute…

• 

  Figure 3. Scatter plots of ICCs for evaluating the reproducibility of radiomic features.The plots illustrate the inter-observer agreement of the extracted high-dimensional features. (A) Features derived from the T2WI sequence, with 950 features retained. (B) Features derived from the CE-T1WI sequence, with 840 features retained. The red horizontal solid line indicates the predetermined reliability threshold of ICC = 0.75. Features plotted above this line were considered highly reproducible and were selected for subsequent dimensionality reduction. ICC, intraclass correlation coefficient; T2WI, T2-weighted imaging; CE-T1WI, contrast-enhanced T1-weighted imaging.

• 

  Figure 4. Feature selection and predictive performance evaluation of the machine learning models. (A) Tuning parameter selection in the LASSO model using 10-fold cross-validation. The y-axis represents the AUC. (B) LASSO coefficient profiles of the selected radiomic features. The vertical dashed line is drawn at the optimal regularization parameter (−log(λ) = 2.525), which minimizes the cross-validation error. (C, D) ROC curves comparing the discrimination performance of the clinical, radiomics, and combined models in the training cohort (C) and the internal testing cohort (D). (E, F) DCA evaluating the clinical utility of the three models in the training cohort (E) and the internal testing cohort (F). The combined model consistently demonstrates a higher net clinical benefit across the threshold probability range of 0.2 to 0.8 compared to single-modality models and default strategies. LASSO, least absolute shrinkage and selection operator; AUC, area under the curve; ROC, receiver operating characteristic; DCA, decision curve analysis.

• 

  Figure 5. Global and local interpretability of the predictive models using SHapley Additive exPlanations (SHAP) analysis. (A) Global feature importance ranking of the radiomics-only model, displaying the mean absolute SHAP values for the top features. (B) SHAP summary beeswarm plot for the radiomics-only model, illustrating the non-linear relationship between feature values (color-coded from low to high) and their impact on model predictions. (C, D) Global feature importance (C) and summary beeswarm plot (D) for the combined model, identifying the length of hospital stay (Hospital_day), hs-CRP, Rad-score, and recurrent attacks as the key predictors. (E, F) Local SHAP waterfall plots for two representative patients. (E) Predicted risk: 41.3%. A short hospital stay (11 days, −0.0259) was counteracted by a high Rad-score (1.82, + 0.0128), recurrent attacks (+0.0166), and elevated hs−CRP (124 mg/L, +0.0349). (F) Predicted risk: 73.4%. A prolonged hospital stay (36 days, +0.302), high Rad−score (1.73, + 0.0320), and elevated hs−CRP (123 mg/L, +0.0313) drove the risk increase.

• 

  Figure 6. Predictive performance and clinical utility of the combined model in the independent external validation cohort. (A) ROC curve demonstrating the robust discriminative ability of the model with an AUC of 0.900. (B) Confusion matrix illustrating the detailed classification results. The confusion matrix demonstrated that 95 of 119 patients were correctly classified, including 71 true negatives and 24 true positives. However, notably, the model exhibited a relatively high false-negative rate, misclassifying 18 out of 42 true cases as low-risk, while maintaining a low false-positive count (n = 6).while (C) DCA indicating the clinical applicability of the combined model. The model yields a greater net clinical benefit compared to the default “treat-all” (solid black line) and “treat-none” (dotted horizontal line) strategies across a broad range of threshold probabilities. AUC, area under the curve; ROC, receiver operating characteristic; DCA, decision curve analysis.