An explainable machine learning framework integrating clinical and lipidomic signatures for early severity stratification of hypertriglyceridemic pancreatitis.

Background: Hypertriglyceridemic pancreatitis (HTGP) is increasingly recognized as a major etiology of acute pancreatitis and is associated with a higher risk of early severe progression, systemic complications, and mortality compared with other etiologies. However, currently available scoring systems demonstrate limited accuracy and poor mechanistic interpretability in HTGP. Aim: To develop and validate an explainable machine learning model for early stratification of severe progression in HTGP and to explore its underlying lipidomic mechanisms. Methods: A retrospective cohort of 288 patients with HTGP admitted between 2020 and 2025 was included. Patients were categorized into mild acute pancreatitis (MAP) and moderately severe/severe acute pancreatitis (MSAP/SAP). Ensemble feature selection integrating eight algorithms was applied to identify robust predictors. Ten machine learning models were constructed and compared. SHAP (SHapley Additive exPlanations) analysis was used to interpret model outputs. Untargeted serum lipidomics combined with KEGG pathway enrichment analysis was performed in 15 matched severe/non-severe AP pairs to explore the biological relevance of identified predictors. Results: Among 288 patients, 150 (52.1%) were classified as MSAP/SAP at hospital admission. Ten predictors associated with inflammatory activation, coagulation dysfunction, endothelial leakage, and systemic injury were identified, including IL-6, D-dimer, PCT, albumin, CRP, SIRS, pleural effusion, and pancreatitis-associated ascitic fluid. Among all candidate algorithms, the Naive Bayes model achieved the best overall discrimination (AUC = 0.841), outperforming APACHE II and modified Marshall scores. Calibration assessment revealed a calibration intercept of -0.0332 and a Brier score of 0.2035; the calibration slope of 0.2009 indicated compression of predicted probabilities, suggesting caution in absolute risk estimation. Decision curve analysis demonstrated positive net clinical benefit across threshold probabilities from 0 to approximately 0.85, with the Naive Bayes model outperforming both benchmark scoring systems across clinically relevant thresholds. SHAP analysis identified IL-6, D-dimer, and PCT as the most influential predictors. Exploratory lipidomic pathway enrichment in a sub-cohort of 15 matched pairs identified alterations in glycerophospholipid metabolism, sphingolipid metabolism, necroptosis, and autophagy pathways, providing hypothesis-generating biological context for the identified clinical predictors. Conclusion: We developed an explainable machine learning framework integrating clinical and exploratory lipidomic evidence for early severity stratification of HTGP at hospital admission. The identified predictors converged on interconnected biological axes involving inflammation, coagulation, and endothelial dysfunction. The lipidomic sub-study provided complementary pathway-level context supporting the biological plausibility of these predictors. As this study is based on a single-center retrospective cohort without external validation, findings should be regarded as preliminary; prospective multicenter validation is required before clinical deployment.