Predicting unplanned CRRT interruptions in critically Ill burn patients using stacked ensemble machine learning framework.

Abstract

Unplanned interruption of continuous renal replacement therapy (CRRT) is a frequent complication in critically ill burn patients that may compromise therapeutic efficacy and clinical outcomes, yet accurate risk identification remains challenging due to complex interactions among patient characteristics, coagulation status, and treatment parameters. In this retrospective cohort study of 666 critically ill burn patients undergoing CRRT (training cohort 2015-2023, n=560; temporal validation cohort 2024-2026, n=106), we identified 12 independently associated variables through multivariate logistic regression and systematically evaluated 47 machine learning algorithms, ultimately constructing a stacking ensemble model integrating bayesglm, fda, knn, and naive_bayes as base learners. In the temporal validation cohort, the ensemble model achieved an AUC of 0.963, accuracy of 0.943, sensitivity of 1.000, specificity of 0.896, and a Kappa of 0.887, with satisfactory calibration and significant clinical net benefit across a wide range of threshold probabilities. Two-level SHAP analysis identified blood flow rate, anticoagulation strategy, hematocrit, filtration fraction, and other CRRT-related parameters as substantial contributors to prediction. This interpretable ensemble model demonstrates robust discrimination, calibration, and clinical utility, and may support early nursing surveillance, individualized CRRT management, and preventive interventions to reduce unplanned interruptions in burn intensive care settings.