Review Identifies Gaps in Surgical ML Risk Prediction

Postoperative complications and mortality remain a substantial global health burden, with early identification of high-risk patients being crucial for preventative care. Machine learning (ML) offers a data-driven approach to model complex clinical patterns using electronic health records (EHRs). However, a comprehensive scoping review of 190 studies on ML for surgical risk stratification and outcome prediction has uncovered significant inconsistencies and limitations in current practices.The review highlighted several critical gaps across the ML workflow. Most studies relied on private, single-center datasets, severely limiting reproducibility and generalizability due to a scarcity of open-access surgical datasets. Reporting of essential preprocessing steps, such as handling missing data, feature selection, and class imbalance, was often incomplete.Furthermore, the research found a predominance of conventional ML models and simple neural networks, with deep learning and multimodal approaches being uncommon. The absence of benchmark datasets and standardized evaluation protocols hindered cross-study comparisons. Only about a third of the studies incorporated explainability methods. These findings underscore the need for more rigorous, reproducible, and clinically meaningful ML development in perioperative care.