J Thorac Cardiovasc Surg. 2025 Dec 5:S0022-5223(25)01041-4. doi: 10.1016/j.jtcvs.2025.11.020. Online ahead of print.
ABSTRACT
OBJECTIVE: Predicting prolonged intensive care unit (ICU) length of stay (LOS) remains challenging, and traditional statistical models often fail to capture patient's complexities. Recent artificial intelligence tools, such as supervised machine learning (ML) algorithms can provide new insights in the field. This study aims to develop and cross-validate a collection of supervised ML models to predict prolonged ICU LOS.
METHODS: Adult patients submitted to cardiac surgery were categorized into short (up to two days) or prolonged (more than two days) ICU LOS in the Cardiothoracic ICU. Information gain analysis identified the most relevant predictors of prolonged ICU LOS to be incorporated in the ML algorithms. Multiple supervised ML approaches were used to predict prolonged ICU LOS, comparing the performance of each model. The clinical applicability of the developed model was assessed with the creation of an interactive web-based application.
RESULTS: Prolonged ICU LOS occurred in 48.8% of the 1387 patients included. The strongest predictors of prolonged ICU LOS were the Sequential Organ Failure Assessment (SOFA) and vasoactive-inotropic (VIS) scores at 24h after surgery, preoperative NT-proBNP and creatinine levels, and cardiopulmonary bypass time. The Random Forest model had better predictive performance (area under the curve 0.804) with 67.8% sensitivity and 81.8% specificity compared to other models.
CONCLUSIONS: Supervised ML models offer a reliable approach to predict prolonged ICU LOS using perioperative data, in patients admitted to a Cardiothoracic ICU. We highlight the potential of clinical applicability of AI-enhanced models with the creation of an interactive widely available web-based application.
PMID:41354172 | DOI:10.1016/j.jtcvs.2025.11.020