Prediction of ICU needs and organ failures in infectious patients using machine learning

This retrospective multicenter cohort study developed predictive models to identify patients at high risk for intensive care and major organ failure, using data from the First Affiliated Hospital of Xiamen University and the MIMIC-III/IV databases. A total of 18,689 infectious patients were analyzed. Nine machine learning algorithms were employed, with CatBoost demonstrating the best performance. The intensive care model achieved an area under the receiver operating characteristic curve (AUC) of 0.956, while organ failure prediction models for renal, heart, respiratory, hepatic, and coagulation systems showed strong accuracy, with most models achieving AUCs over 0.8. Random Forest performed best for predicting respiratory and renal failures, with AUCs of 0.677 and 0.836, respectively. Model interpretability was supported by real clinical cases, where abnormal indicators aligned with key predictive features. These models offer valuable insights for early decision-making, enhancing resource allocation and patient outcomes.