Can Machine Learning Methods Improve the Prediction of Postoperative ICU Requirements? Real-World Evidence and Practical Implications

Accurate prediction of postoperative care requirements following elective surgical procedures is of central importance from both medical and organizational perspectives. Although machine learning (ML) methods have been developed to support such predictions, clinical decisions regarding postoperative care levels are still predominantly based on the individual judgment of surgeons and anesthesiologists. Methods: This retrospective, single-center observational study was conducted at the University Hospital Augsburg and includes 35,488 elective surgical cases documented between August 1, 2023, and January 31, 2025. For each case, the preoperative care level prediction made by surgical and anesthesiology teams was compared to the actual post-operative care provided. The predictive performance of clinical assessments was further benchmarked against ML-based approaches described in the current literature. Results: While overall prediction accuracy was high (surgical assessments: 91.17%; anesthesiology assessments: 87.12%), the sensitivity for identifying patients requiring intensive care remained markedly lower than that reported for ML-based models. This discrepancy was particularly pronounced for patients ultimately admitted to the ICU postoperatively. Despite good overall accuracy, clinical predictions by physicians show limited sensitivity in identifying high-risk patients in need of postoperative intensive care. Conclusion: This study presents, for the first time, a data-driven framework to support and enhance current preoperative decision-making through the integration of ML techniques.