Machine learning in critical care: Improving prediction of mortality and intensive care unit stay after cardiac surgery.

Objective: The European System for Cardiac Operation Risk Evaluation II (EuroSCORE II) is a multifactorial tool that assesses postoperative mortality risk after cardiac surgery but lacks several patient-specific and procedure-related factors. This study evaluated machine learning (ML) models against the EuroSCORE II for predicting in-hospital mortality and prolonged intensive care unit (ICU) stay after cardiac surgery. Methods: In this retrospective single-center study, data from 5606 adult patients undergoing primary open-heart surgery were analyzed. Six ML models were compared with EuroSCORE II using area under the receiver operating characteristic (AUROC) analysis for mortality and ICU length-of-stay predictions using a specified set of pre- and perioperative parameters. Feature importance for each prediction task was assessed using the random forest classifier. Results: Coronary artery bypass grafting was the most common procedure (50.2%), followed by aortic valve (20.7%), mitral valve (16.4%), thoracic aorta (8.8%), tricuspid valve (0.7%), and other procedures (3.1%). EuroSCORE II predicted a 2.26% mortality rate, whereas the actual in-hospital mortality was 1.5% (83 patients). The Gaussian Process (GP) classifier surpassed EuroSCORE II for prediction of in-hospital mortality with AUROC values of 0.877 versus 0.851 using preoperative data, reaching 0.886 when perioperative data were included. For prolonged ICU stay prediction, the GP classifier achieved AUROCs of 0.725 (preoperative data) and 0.749 (perioperative data). Conclusions: EuroSCORE II overestimated mortality despite a robust performance. Among tested ML models, the GP classifier demonstrated superior accuracy using preoperative data, further enhanced by perioperative parameters. Our results highlight ML's potential for improving risk assessment by its integration into next-generation cardiac surgery risk assessment tools.