Using machine learning and preoperative data to predict 30-day mortality after non-cardiac major surgeries in older patients

Postoperative mortality remains a critical concern, especially among older patients undergoing non-cardiac major surgeries. With over 300 million surgeries performed worldwide annually, a significant proportion involves elderly patients, making early and precise risk stratification essential. Traditional assessment tools often require complex calculations or rely on preoperative medical coding systems that vary across regions, limiting their widespread applicability. Machine learning (ML) presents a promising solution by leveraging readily available patient data to enhance near-real-time preoperative risk assessment, adapting to the specific needs of each hospital. This retrospective study utilized data from Taichung Veterans General Hospital, including patients aged 65 and older who underwent non-cardiac major surgery between January 2017 and April 2019. Preoperative data, including demographic characteristics, laboratory values, comorbidities, surgical risk classification, and the Age-Adjusted Charlson Comorbidity Index (ACCI), were extracted from electronic medical records. The primary outcome was 30-day postoperative mortality. We employed logistic regression, random forest (RF), XGBoost (XGB), gradient boosting machine (GBM, implemented via scikit-learn), and support vector classification (SVC), to develop predictive models. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC). A total of 9,422 patients were included, with 1.6% ( n = 147) experiencing 30-day mortality. XGBoost demonstrated good predictive performance, achieving an AUROC of 0.835 when integrating basic features, laboratory data, comorbidities, surgical risk, and ACCI. Feature importance analysis identified the highest-ranking feature was hematocrit, followed by prothrombin time, sodium, ACCI, and ASA classification. A reduced model using the top 15 features maintained comparable performance (AUROC = 0.819). Our study demonstrates that ML models, particularly XGBoost, can effectively predict 30-day mortality in elderly patients undergoing non-cardiac major surgery. The model’s integration of preoperative clinical and laboratory data enhances its practical application for near-real-time risk stratification. Implementing ML-based predictive tools in preoperative assessments may facilitate early intervention and improve patient outcomes.