Key result
Machine learning models accurately predict postoperative mortality with validation AUROCs reaching ~0.90.
Why the study?
Existing surgical risk prediction tools are often limited by short-term outcomes, prompting the development and validation of perioperative and preoperative models for predicting mortality and major complications.
Cohort (n=248,346)
Yes
Answers are grounded in this paper and cite their sources.
Machine learning-based perioperative and preoperative surgical risk prediction models developed from 11 years of population-based data in Hong Kong demonstrated high discrimination (AUROC 0.89-0.91) for predicting 30-, 60-, and 90-day mortality and major complications.
Background Existing surgical risk prediction tools are often limited by short-term outcomes. Our study aimed to develop and validate perioperative models and preoperative models for predicting mortality and major complications using surgical data from Hong Kong. Methods This population-based retrospective cohort study analysed 11 years of surgical data (July 2014–June 2024) collected from 17 public hospitals in Hong Kong. Of 124 clinically relevant candidate predictors, 67 predictors were retained for the perioperative full models. To improve clinical usability and facilitate preoperative risk assessment, reduced preoperative models were trained using 13 predictors. Separate models were developed for emergency and elective surgery. Primary outcomes were 30-day, 60-day, and 90-day all-cause mortality in emergency and elective surgery cohorts. Secondary outcomes included 30-day major surgical-related and medical-related complications. Multiple risk prediction models were evaluated using training data from 2014 to 2021 with 5-fold stratified cross-validation, while temporal external validation was performed using data from 2022 to 2024. Performance was assessed using the area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), Brier score, calibration intercept and slope, integrated calibration index (ICI), E50, and E90. SHAP was used for model interpretation. Models showing poor calibration underwent logistic recalibration using five-fold out-of-fold development predictions. Findings 43,601 subjects were included in the training set and 16,143 in the external validation set for emergency surgeries, whereas 151,016 subjects were used for training and 37,586 for external validation for elective surgeries. In temporal validation, the perioperative full models achieved AUROCs of 0.89, 0.89, and 0.89 for 30-day, 60-day, and 90-day mortality after emergency surgery and 0.91, 0.90, and 0.90 after elective surgery. Corresponding AUPRCs were 0.27, 0.36, and 0.41 for emergency surgery and 0.05, 0.10, and 0.13 for elective surgery. AUROCs for medical- and surgery-related complications were 0.88 and 0.78 in emergency surgery and 0.87 and 0.80 in elective surgery. The 13-predictor preoperative models retained useful discrimination, although performance was lower than that of the perioperative full models. Calibration assessment indicated variable agreement between predicted and observed risks, supporting the need for recalibration before clinical implementation. Eight complication-specific models showed validation AUROCs ranging from 0.75 to 0.94. SHAP analysis revealed clinically plausible risk factors, such as higher ASA status, abnormal laboratory results (such as renal dysfunction and low albumin), and surgical complexity. Logistic recalibration improved Brier scores and calibration indices across all full and reduced models without changing AUROC or AUPRC. Interpretation We have developed and validated comprehensive surgical risk prediction models with excellent discrimination and robust performance using 11 years of population-based data from Hong Kong. The identified predictors were clinically interpretable, which can serve as a data-driven tool to assist clinicians and patients for surgical risk assessments.
No takes yet. Share an insight, caveat, or question.
Wei et al. (2026) conducted a cohort in Patients undergoing emergency or elective surgery (n=248,346). Machine learning perioperative (67-predictor) and preoperative (13-predictor) risk prediction models vs. Reduced 13-predictor preoperative models and temporal external validation cohorts was evaluated on 30-day, 60-day, and 90-day all-cause mortality in emergency and elective surgery cohorts (AUROC 0.89 (emergency 30/60/90-day) and 0.91, 0.90, 0.90 (elective 30/60/90-day)). Perioperative machine learning models predicted 30-, 60-, and 90-day postoperative mortality with validation AUROCs of 0.89 for emergency surgery and 0.90–0.91 for elective surgery.
Synapse has enriched 10 related papers. Consider them for comparative context: