Construction and validation of risk prediction models for renal replacement therapy in patients with acute pancreatitis.

Background: Renal replacement therapy (RRT) plays a crucial role in managing acute pancreatitis (AP). This study aimed to develop and evaluate predictive models for determining the need for RRT among patients with AP in the intensive care unit (ICU).

Methods: A retrospective selection of patients with AP was made from the Medical Information Mart for Intensive Care IV (MIMIC-IV, version V2.0). The cohort was randomly divided into a training set (447 patients) and a validation set (150 patients). The least absolute shrinkage and selection operator (LASSO) regression cross-validation method was utilized to identify key features for model construction. Using these features, four machine learning (ML) algorithms were developed. The optimal model was visualized and clarified using SHapley Additive exPlanations (SHAP) and presented as a nomogram.

Results: The mean age of the cohort was 59.17 years, with an average Acute Physiology and Chronic Health Evaluation II (APACHE II) score of 17.55. Acute kidney injury (AKI) was observed in 52.43% of patients with AP, and 9.05% required RRT. After feature selection, four of 41 clinical factors were ultimately chosen for use in model construction. The Lasso-Logistic Regression (Lasso-LR) model showed a high discriminative ability to predict RRT risk in patients with AP, with an area under the receiver operating characteristic (AUROC) of 0.955 (95% CI 0.924-0.987) in the training set. In the validation set, it maintained its discriminative performance, achieving an AUROC of 0.985 (95% CI 0.970-1.000). Calibration curves indicated an excellent fit in both sets (Brier scores: 0.039 and 0.032, respectively), suggesting high consistency. Decision curve analysis (DCA) highlighted the Lasso-LR model's significant clinical utility in predicting RRT likelihood in patients with AP.

Conclusions: Developed via the LASSO regression cross-validation method, the Lasso-LR model significantly excels in predicting the requirement for RRT in patients with AP, demonstrating its potential for clinical application.