Utilizing deep learning models on separately defined risk cohorts significantly improved predictive accuracy for postoperative complications. Training on high-risk patients yielded better area under the precision-recall curve for predicting in-hospital mortality, acute kidney injury, and prolonged ICU stays. This tailored approach notably enhanced F1 scores for various complications, suggesting that risk-specific training could address class imbalance issues in surgical risk prediction, thereby enabling better clinical decision-making in surgical settings.
Journal Article by Balch JA, Ruppert MM (…) Loftus TJ et 8 al. in JAMA Surg