A large-scale, population-based cohort study conducted across 17 public hospitals in Hong Kong has demonstrated the efficacy of machine learning–based predictive modeling for estimating postoperative mortality and major complications over extended time horizons. By analyzing eleven years of territory-wide surgical data, researchers developed robust ensemble models capable of predicting 30-, 60-, and 90-day all-cause mortality as well as 30-day major morbidities for both emergency and elective surgical cohorts. The study, which included data collected from July 2014 to June 2024, utilized a sophisticated blended ensemble approach, achieving consistent and high discriminative performance with areas under the receiver operating characteristic curve (AUROC) reaching approximately 0.9 for mortality outcomes across all three timeframes. This advancement addresses a critical gap in current surgical risk assessment, as most traditional clinical tools are restricted to 30-day outcomes and lack the capability to provide extended postoperative monitoring or comprehensive morbidity prediction.
Surgical intervention remains an essential, high-volume component of modern healthcare systems globally, with approximately 313 million procedures performed annually. Despite the clinical necessity of these interventions, the burden of adverse postoperative outcomes is substantial, contributing to millions of deaths within 30 days following surgery each year. While standardized risk prediction tools such as the American College of Surgeons (ACS) National Surgical Quality Improvement Program (NSQIP) calculator, the Surgical Risk Preoperative Assessment System (SURPAS), and the Surgical Outcome Risk Tool (SORT) have historically supported patient counseling and clinical decision-making, they often rely on a limited number of variables and a constrained prediction window. Extending the prediction horizon to 60 and 90 days offers a more comprehensive assessment of surgical quality, as these metrics are less susceptible to institutional variations in hospital discharge practices and provide a clearer picture of delayed postoperative risk. By integrating heterogeneous surgical data into multi-variable machine learning frameworks, this research aimed to provide a more scalable and accurate tool for contemporary surgical risk management.
The study population was derived from the Surgical Outcomes Monitoring and Improvement Program (SOMIP), a quality assurance initiative in Hong Kong that systematically audits patient-level perioperative data. Researchers included 43,601 subjects in the training set and 16,143 in the external validation set for emergency surgeries, and 151,016 training subjects compared to 37,586 external validation subjects for elective procedures. To capture the full breadth of preoperative and intraoperative clinical influences, 67 predictors were extracted from clinical information systems, encompassing demographic information, extensive preoperative variables, and laboratory test results. The research team distinguished between emergency procedures—defined by immediate threats to life, limb, or organ—and elective procedures conducted under stable clinical conditions, recognizing that these two surgical cohorts present fundamentally different baseline risk profiles and clinical trajectories.
The modeling process involved a rigorous evaluation of various prediction algorithms, including Gradient Boosting, AdaBoost, CatBoost, Random Forest, and other tree-based classifiers. These models were trained on surgical records from 2014 through 2021, while temporal validation was conducted using the subsequent 2022 to 2024 datasets. The final predictive models were constructed as a stacked ensemble, which integrated the complementary predictive strengths of multiple individual classifiers to optimize overall performance. Model discrimination was evaluated via AUROC and the area under the precision-recall curve (AUPRC), with calibration assessed through the Brier score, calibration slope, and the integrated calibration index (ICI). This comprehensive validation framework ensured that the models remained reliable when applied to unseen temporal data, reflecting realistic clinical deployment scenarios.
Results from the validation cohorts confirmed the strong performance of the blended ensemble model. For mortality prediction in emergency surgery, the model achieved an AUC of 0.89 for 30-, 60-, and 90-day mortality. Elective surgery mortality prediction yielded even higher discrimination, with AUCs of 0.91, 0.90, and 0.90 for the 30-, 60-, and 90-day timeframes, respectively. While predictive discrimination for medical-related complications was excellent—achieving an AUC of 0.88 in emergency surgery and 0.87 in elective cohorts—performance for surgery-related complications was comparatively lower at approximately 0.78 to 0.80. The research team noted that AUPRC values were more modest for rare outcomes, particularly within the elective surgery group, underscoring the nuance that while models excel at ranking patients by relative risk, absolute risk calibration requires careful attention before clinical implementation.
To enhance the interpretability of these complex machine learning models, the researchers utilized Shapley Additive Explanations (SHAP), a game-theoretic framework that quantifies the contribution of individual features to each specific prediction. These analyses revealed that serum albumin level was consistently the most influential predictor for mortality across both emergency and elective populations, reinforcing its role as a vital marker of physiological reserve. For surgery-related complications, predictors such as surgical complexity score, preoperative dyspnoea at rest, and elevated serum urea were identified as major contributors, highlighting the intersection of procedural difficulty and patient vulnerability. These SHAP findings provided clinically plausible explanations for model outputs, bridge the gap between opaque algorithmic predictions and actionable clinical insights for healthcare providers.
A secondary goal of the study was to improve the preoperative utility of the models through feature reduction. By applying various feature-selection methods across model families, the researchers identified a subset of 13 key variables that retained much of the predictive performance of the full 67-predictor models. These reduced preoperative models, which include factors such as age-related comorbidity burden, ASA physical status, specific laboratory abnormalities, and procedural complexity, achieved validation AUCs of 0.86 to 0.87 for mortality in emergency surgery and 0.84 to 0.86 in elective settings. The conciseness of the reduced models supports their potential for rapid preoperative planning and individual patient counseling, offering a streamlined interface for clinical environments where full perioperative datasets may not be immediately available.
Although the models demonstrated strong initial performance, the study highlighted the importance of recalibration for maintaining the accuracy of absolute risk estimates. Logistic recalibration was performed to improve agreement between predicted and observed risks, especially in the elective cohort where original models showed signs of overprediction. Following this procedure, validation calibration slopes for both full and reduced models moved close to the ideal value of 1.00, significantly reducing the integrated calibration index for all investigated outcomes. This recalibration step underscores the iterative nature of machine learning deployment, demonstrating that high discriminative power does not automatically equate to perfect absolute risk accuracy and necessitates rigorous temporal adjustments.
Several limitations inherent to the study design must be considered. The research was confined to public hospitals in Hong Kong, and while temporal validation was performed using data from 2022 to 2024, the findings may not be fully generalizable to private healthcare systems or international populations. Furthermore, the dataset lacked certain nuanced variables such as detailed frailty indices, socioeconomic markers, and patient-reported outcomes, which could potentially improve risk stratification. The study also relied on retrospective observational data, meaning that the identified associations between variables like serum albumin and mortality are associative rather than strictly causal. Consequently, the research team emphasizes that these models do not currently establish defined thresholds for clinical intervention and should be subjected to prospective clinical impact evaluations before widespread institutional adoption.
The study concludes that robust, data-driven surgical risk prediction is feasible and capable of capturing outcomes beyond the standard 30-day window. By providing tailored predictions for 30-, 60-, and 90-day intervals, these machine learning models offer a more granular view of the surgical landscape, which may facilitate better resource allocation and individualized care planning. The findings suggest that the integration of demographic, clinical, and laboratory measures allows for a comprehensive assessment of patient physiological reserve. While further research is needed to evaluate real-world utility and potential impacts on decision-making processes, the development of these interpretable, highly accurate tools represents a significant step toward improving surgical safety and quality management in the face of growing global demand for surgical services.
Subject of Research: Machine learning-based prediction of postoperative mortality and morbidity.
Article Title: Machine learning–based prediction of 30-, 60-, and 90-day postoperative mortality and major complications: an 11-year population-based cohort study in Hong Kong
Article References: Wei, Y., Zhang, W., Lin, G., Li, C., Tung Lam, T. Y., Hung, C. T., Leung, E., Yeoh, E. K., Chong, K. C., Cheung, Y. S., & San Lai, P. B. (2026). Machine learning–based prediction of 30-, 60-, and 90-day postoperative mortality and major complications: an 11-year population-based cohort study in Hong Kong. The Lancet Regional Health – Western Pacific, 76, Article 102027. https://doi.org/10.1016/j.lanwpc.2026.102027
Image Credits: AI Generated
DOI: 10.1016/j.lanwpc.2026.102027
Keywords: surgical risk prediction, machine learning, postoperative mortality, postoperative complications, retrospective cohort study, ensemble learning, SHAP analysis, emergency surgery, elective surgery, clinical decision support, surgical quality improvement, Hong Kong healthcare, perioperative care, mortality prediction, morbidity prediction, predictive modeling, healthcare analytics, patient stratification
News Source: Teresa Odom. (November 17, 2026). Machine Learning Models Predict Long-Term Surgical Mortality Across Emergency and Elective Patients. Scienmag.



