Machine learning algorithms to predict acute respiratory distress syndrome in critically ill patients with acute pancreatitis from the MIMIC-IV database

Acute pancreatitis (AP) frequently leads to acute respiratory distress syndrome (ARDS), markedly increasing mortality among critically ill patients. Traditional prediction models often fail to accurately identify early ARDS risk. This study aims to develop and validate novel machine-learning models based on clinical features for early ARDS prediction in critically ill AP patients. Data from patients with AP were extracted from the Medical Information Mart for Intensive Care-IV 2.2 (MIMIC-IV v2.2) database. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO) regression. Seven machine learning algorithms were applied for model development. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, precision, F1 score, and Brier score. The SHapley Additive exPlanations (SHAP) method was used to provide visual interpretability of the best model. Furthermore, calibration and decision curves were employed to assess the prediction results. A total of 795 critically ill patients with AP were eligible for participation, of which 140 (17.6%) developed ARDS during intensive care units (ICU) hospitalization. Ten variables were selected to construct the machine learning models. Among the seven models developed, the random forests (RF) model demonstrated competitive and well-balanced predictive performance, achieving an AUC of 0.804 (95% CI: 0.725–0.878) and tying with XGBoost for the lowest Brier score of 0.121 in the testing set. SHAP analysis revealed several variables important for predicting ARDS: ventilation, the sequential organ failure assessment score, heart rate, albumin, respiratory rate, and the simplified acute physiology score II. SHAP force plots illustrated how the RF model visualizes the individualized prediction of ARDS risk in AP patients. Furthermore, the calibration curve showed good concordance between the predicted and observed values, and decision curve analyses indicated that the model was useful for supporting clinical decisions. Machine learning models, particularly the RF model, show promise in predicting delayed ARDS in AP patients; however, external validation in independent multicenter cohorts is required before clinical implementation.