Machine learning model for early prediction of acute respiratory failure in acute pancreatitis: Multicenter validation

Acute respiratory failure (ARF) is a common organ failure in acute pancreatitis (AP) with high mortality. This study used machine learning to predict ARF risk in AP patients through a retrospective, multicenter cohort analysis involving MIMIC-IV and Chinese hospital data (2010–2025). Variable selection combined SHapley Additive exPlanations values and Lasso regression, employing five machine learning algorithms for modeling. The dataset identified nine most influential features, including temperature, SpO 2 , and seven serum indicators, with the logistic regression (LR) model achieving the highest areas under the curve (AUCs) (0.9048 internal, 0.8717 external, and 0.8557 cross-specialty) and top sensitivity (0.8947 internal, 0.7361 external, and 0.9 cross-specialty). The LR model proved to be the most effective classifier for early ARF prediction in AP. A cross-specialty validation of pregnancy patients delineated the model's boundary condition. An online computing platform for this LR model was publicly available and free to use.