Accurate prediction of intensive care unit (ICU) admission and length of stay (LOS) after major surgery is essential for optimizing patient outcomes and healthcare resources. Factors such as age, BMI, comorbidities, and perioperative complications significantly influence ICU admissions and LOS. Machine learning methods have been increasingly utilized to predict these outcomes, but their clinical utility beyond traditional metrics remains underexplored.

This study examined a sub-cohort of 6043 patients who underwent general anesthesia at Seoul National University Hospital from August 2016 to June 2017. Various prediction models, including logistic regression and random forest, were developed for ICU admission and different LOS thresholds, e.g., a LOS of more than a week. Clinical utility was evaluated using decision curve analysis (DCA) across predefined risk preferences.

Among patients studied, 19.8 % were admitted to the ICU, with 1.4 % staying longer than a week. Prediction models demonstrated high discrimination (AUROC 0.93 to 0.96) and good calibration for ICU admission and short LOS. DCA revealed that intraoperative data provided the greatest decision-related benefit for predicting ICU admission, while preoperative data became more important for predicting longer LOS.

Intraoperative data are crucial for immediate postoperative decisions, while preoperative data are essential for extended LOS predictions. These findings highlight the need for a comprehensive risk assessment approach in perioperative care, utilizing both preoperative and intraoperative information to enhance clinical decision-making and resource allocation.