The role of artificial intelligence in determining the severity of acute pancreatitis

This study aimed to evaluate the diagnostic and prognostic value of artificial intelligence (AI)–based machine learning models in predicting disease severity and major clinical outcomes in patients with acute pancreatitis and to compare their performance with classical clinical scoring systems, namely the Ranson (0-hour) and BISAP scores. In this retrospective observational study, patients diagnosed with acute pancreatitis were classified according to the presence of organ failure, development of pancreatic necrosis, and disease severity based on the Revised Atlanta classification. Clinical, laboratory, and radiological data were statistically analyzed, and AI models based on Random Forest, Gradient Boosting, and Ensemble Learning were developed and compared with conventional scoring systems. The results demonstrated that patients who developed organ failure were older and had significantly longer hospital stays and higher intensive care unit requirements. Levels of 48-hour C-reactive protein (CRP), leukocyte count, creatinine, urea, BUN, lactate, and lower calcium levels were significantly associated with disease severity and organ failure. Among comorbidities, hypertension and coronary artery disease were associated with the development of organ failure. Etiological analysis revealed that non-biliary etiologies were more frequently observed in patients with pancreatic necrosis and in the moderately severe to severe disease groups. Regarding classical scoring systems, a BISAP score ≥3 significantly predicted both organ failure and pancreatic necrosis, whereas the Ranson (0-hour) score did not demonstrate meaningful discriminatory power in the early phase. In contrast, AI-based models showed markedly superior predictive performance compared with classical scoring systems in predicting organ failure, pancreatic necrosis, and severe acute pancreatitis according to the Revised Atlanta classification (AUC ≈ 0.999–1.000). However, the single-center retrospective design and the contribution of time dependent variables, particularly 48-hour CRP, raise concerns regarding potential overfitting. Therefore, multicenter prospective studies with external validation are required to confirm the generalizability of these findings. In conclusion, AI-based machine learning models demonstrate higher predictive accuracy than conventional scoring systems in assessing disease severity and clinical outcomes in acute pancreatitis. Nevertheless, rather than replacing traditional clinical assessment tools, these models should be integrated as complementary decision-support systems to enhance early risk stratification and individualized patient management. Keywords: Acute pancreatitis, disease severity, artificial intelligence, machine learning, organ failure, pancreatic necrosis, Revised Atlanta classification, BISAP score, Ranson score