Answers are grounded in this paper and cite their sources.
Background Capillary leak syndrome (CLS) is a clinically relevant complication of acute pancreatitis (AP). However, reliable approaches for predicting CLS in AP remain limited, and the association between CLS and severe AP (SAP) remains unclear. Methods We retrospectively enrolled 337 patients with AP (89 with CLS and 248 without CLS). Forty-three candidate demographic and laboratory variables were collected. Six machine learning (ML) models were trained using nested cross-validation with LASSO-based feature selection. Model performance was assessed using the ROC-AUC, PR-AUC, Brier score, calibration, decision curve analysis, and model stability, and a composite score was computed for overall evaluation. The optimal model was interpreted using SHAP. We assessed the association between model-derived CLS risk and SAP using group comparison, discrimination, dose–response, and logistic regression analyses. Results Among the enrolled patients, 89 had CLS. XGBoost was selected as the optimal model, with an ROC-AUC of 0.918 (95% confidence interval [CI]: 0.882–0.949), a PR-AUC of 0.784 (95% CI: 0.682–0.872), and the lowest Brier score (0.093, 95% CI: 0.070–0.117). SHAP analysis identified HGB, NEUT%, UA, GGT, TC, and HDL-C as predictors of CLS. CLS risk was significantly higher in SAP patients and differed significantly across the four subgroups stratified by SAP and CLS status ( P < 0.001), with a higher risk in CLS-positive subgroups. The proportion of SAP increased from 11.8% in Q1 to 63.1% in Q4 ( P < 0.0001), and CLS risk discriminated SAP with an AUC of 0.766 (95% CI: 0.716–0.816). Each 0.1 increase in CLS risk was associated with higher odds of SAP (adjusted odds ratio: 1.23; 95% CI: 1.15–1.33). Conclusion These findings suggest that explainable ML-based CLS risk estimates may help identify AP patients with clinical features associated with CLS and provide complementary information regarding SAP risk stratification.
No takes yet. Share an insight, caveat, or question.
Ji et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: