Predicting KRT or Death in Critically Ill Patients with Rhabdomyolysis Using Machine Learning

BACKGROUND: Rhabdomyolysis is associated with outcomes ranging from muscle injury to AKI, KRT, and death. Early identification of patients at highest risk of severe outcomes may help inform triage, monitoring, and nephrology consultation. We aimed to develop and externally validate a machine-learning model to predict in-hospital death or KRT in patients with rhabdomyolysis. METHODS: We analyzed two independent United States intensive care unit databases: the Philips eICU Collaborative Research Database (eICU) and the Medical Information Mart for Intensive Care IV (MIMIC). Adults with creatine kinase (CK) greater than 5,000 IU/L within 72 hours of admission and length of stay greater than one day were included. Patients receiving KRT before admission or with CK elevation due to myocardial infarction were excluded. The eICU cohort was randomly divided into training and internal test sets to develop a random forest model predicting a composite outcome of in-hospital death or KRT. External validation was performed in the MIMIC cohort. Model performance was compared with the McMahon score. RESULTS: A total of 753 and admissions in eICU and 768 in MIMIC were included (30% female, mean age 50 years). The composite outcome occurred in 17% and 10%, respectively. The final model incorporated 15 routinely available variables. In external validation, the model demonstrated an area under the receiver operating characteristic curve of 0.89, sensitivity of 80%, specificity of 84%, and a positive likelihood ratio of 5.01. The model showed better discrimination and greater decision-analytic benefit than the McMahon score. Creatine kinase alone showed poor discrimination for the composite outcome. CONCLUSIONS: A machine-learning model using routinely available clinical data predicted in-hospital death or acute KRT in critically ill patients with rhabdomyolysis and outperformed an established risk score in two intensive care unit cohorts. These findings support further prospective validation before clinical implementation.