Cytokine network analysis and machine learning-based stratification of disease severity in acute pancreatitis

Acute pancreatitis (AP) is an acute abdominal disease characterized by local pancreatic inflammation and complex immune dysregulation. Given the central role of inflammatory cytokines in AP pathophysiology and their potential association with disease severity, we compared plasma cytokine profiles among patients with mild, moderately severe, and severe AP. Correlation patterns among cytokines were further characterized using correlation network analysis, and machine learning approaches were applied for exploratory severity stratification. In this study, we retrospectively included 220 patients with acute pancreatitis (mild, moderately severe, or severe) and retrieved plasma cytokine measurements of eight cytokines (IL-2, IL-4, IL-6, IL-10, IL-12p70, IL-17, IFN-γ, and TNF-α). Correlation network analysis was performed to characterize severity-associated cytokine correlation patterns. Two exploratory machine learning approaches were applied, including LASSO-based multinomial logistic regression for cytokine feature selection and random forest analysis for severe versus non-severe AP stratification. Model performance was evaluated using repeated nested cross-validation, ROC-based discrimination metrics, and variable importance analysis. Cytokine network analysis showed only descriptive differences in correlation patterns across AP severity groups; however, no between-group difference in any individual cytokine–cytokine correlation remained statistically significant after false discovery rate (FDR) correction. Cytokine distributions showed partial overlap across groups, indicating limited separability using single biomarkers. LASSO regression identified a subset of cytokines associated with severity, while random forest modeling demonstrated moderate discriminative ability for severe disease classification after repeated nested cross-validation, with high specificity but limited sensitivity. Feature selection results differed between analytical approaches, but IL-6 and IL-10 were consistently highlighted as informative cytokines. These findings suggest that cytokine-based exploratory modeling may provide complementary information for AP severity stratification. This study provides a preliminary framework for investigating cytokine-based severity stratification and supports future validation studies in larger prospective cohorts.