Machine learning–random survival forest prediction model for predicting short-term recurrence after first-episode acute pancreatitis

Abstract Background This study aimed to construct and assess a machine learning–random survival forest prediction model for predicting short-term recurrence after first-episode acute pancreatitis (AP). Methods A total of 568 patients with AP admitted to Cangzhou Central Hospital from October 2018 to December 2025 were retrospectively analyzed and divided into the recurrent AP group (RAP) and non-RAP group according to whether recurrence occurred within 1 year. In the machine learning section, the patients were randomly divided into a training cohort ( n = 398) and an internal validation cohort ( n = 170) at a 7:3 ratio. An external validation cohort of 77 patients from another hospital was used for model evaluation. Cox regression analysis was applied to screen the predictive factors of recurrence, and a random survival forest model was constructed. The performance of the random survival forest prediction model was assessed using the receiver operating characteristic curve, calibration curve, decision curve analysis (DCA), and Shapley Additive exPlanations (SHAP). Results Approximately 14.1% of the first-episode AP cases (80/568) recurred. Cox analysis identified alcohol drinking (HR = 2.35, 95% CI = 1.37–4.05, P = 0.002), triglyceride (HR = 1.13, 95% CI = 1.10–1.16, P < 0.001), and pancreatic necrosis (HR = 5.34, 95% CI = 2.11–13.54, P < 0.001) as independent predictive factors of RAP. The nomogram prediction model achieved an area under the curve of 0.834 (95% CI: 0.765–0.891) in the training cohort, 0.794 (95% CI: 0.677–0.903) in the internal validation cohort, and 0.794 (95% CI: 0.657–0.909) in the external validation cohort. The top 3 variables in terms of importance were ranked as follows: triglyceride, alcohol drinking, and pancreatic necrosis. The random survival forest prediction model achieved an area under the curve of 0.947 (95% CI: 0.920–0.971) in the training cohort, 0.832 (95% CI: 0.711–0.931) in the internal validation cohort, and 0.815 (95% CI: 0.687–0.906) in the external validation cohort. The calibration curve showed that the random survival forest model had good predictive ability, and DCA confirmed that the model had strong clinical utility. SHAP analysis identified triglyceride as the most critical predictor, with secondary contributions from alcohol drinking and pancreatic necrosis. Conclusion Random survival forest models are insightful and useful tools for predicting short-term recurrence for patients with first-episode AP. Clinical trail Not applicable.