Construction of a predictive model for severe hypertriglyceridemia-associated acute pancreatitis using machine learning

OBJECTIVE: To identify critical risk factors distinguishing severe hypertriglyceridemia-associated acute pancreatitis (SAP-HTG) from its non-severe form (N-SAP-HTG). Machine learning techniques were used to develop predictive models and compare them to conventional scoring systems, aiming to enhance early diagnosis and risk stratification of SAP-HTG. METHODS: A retrospective analysis was conducted on 514 patients with acute pancreatitis admitted to HanZhong Central Hospital between August 2018 and June 2024, including 90 SAP-HTG and 424 N-SAP-HTG cases. Key baseline characteristics, scoring indices (APACHE II, Ranson), and laboratory data (fasting blood glucose (FBG), triglycerides (TG), C-reactive protein (CRP)) were collected. LASSO regression was used to identify key predictors, and multivariate logistic regression was applied to assess their associations. Four predictive models - logistic regression, random forest (RF), support vector machine (SVM), and XGBoost - were developed. Model performance was evaluated using the confusion matrix, receiver operating characteristic (ROC) curves, area under the curve (AUC), and SHAP analysis, with comparisons to APACHE II and Ranson scores. Statistical analyses were conducted with SPSS 26.0 and R 4.3.3. RESULTS: and age were protective (OR = 0.258 and 0.290, respectively). Among the models, XGBoost and RF achieved the highest discriminatory power, with AUCs of 0.959 and 0.955, surpassing logistic regression (0.924), SVM (0.926), and traditional scoring systems (P < 0.002). SHAP analysis revealed CRP, TG, and TyG index as the most influential factors. CONCLUSION: Machine learning models effectively identified SAP-HTG risk factors, with XGBoost showing superior performance over conventional scoring systems. These models provide a valuable tool for early diagnosis and risk stratification of SAP-HTG in clinical settings.