JMIR Publications

ABSTRACT

Background:

Postoperative complications after noncardiac surgery under general anesthesia remain a major cause of morbidity, mortality, and resource-intensive care. Conventional risk assessment tools, which rely largely on preoperative information, often fail to capture perioperative physiological and procedural factors that influence postoperative risk.

Objective:

This retrospective multicenter observational study aimed to develop and validate electronic health record (EHR)–based machine learning models for predicting major postoperative complications and to evaluate their discrimination, calibration, and clinical utility.

Methods:

We analyzed data from 35,702 adult patients who underwent noncardiac surgery under general anesthesia at three tertiary hospitals between January 2021 and March 2025. Routinely available preoperative, intraoperative, and early postoperative variables were extracted from Diagnosis Procedure Combination claims, EHRs, and anesthesia information systems. Major postoperative complications within 30 days—including mortality, unplanned reoperation, prolonged mechanical ventilation, new renal replacement therapy, and postoperative infection—were defined as a composite outcome. Logistic regression, extreme gradient boosting, and light gradient boosting machine (LGBM) models were developed with hyperparameter optimization and isotonic calibration. Model performance was assessed using the area under the precision–recall curve (PR-AUC), Brier score, decision curve analysis, and internal–external cross-validation across institutions.

Results:

Postoperative complications occurred in 3,884 of 35,702 patients (10.9%). Models incorporating perioperative variables demonstrated superior discrimination compared with preoperative-only models (PR-AUC 0.43–0.48 vs 0.39) and improved calibration (Brier score 0.076 vs 0.089). Decision curve analysis showed consistently greater net benefit for perioperative models across clinically relevant risk thresholds (0.05–0.30). Internal–external cross-validation confirmed stable discrimination across sites, although calibration varied between institutions, indicating the need for site-specific recalibration.

Conclusions:

Machine learning models using routinely available perioperative EHR data provide clinically meaningful improvement in predicting major postoperative complications after general anesthesia. Because all predictors are routinely documented in perioperative EHR systems, these models are designed to be readily implementable as early postoperative clinical decision support tools. Gradient boosting models offer complementary strengths, with well-calibrated models supporting probability-based intervention thresholds and highly discriminative models aiding high-risk patient prioritization. Future studies should focus on prospective validation, local recalibration, and postdeployment performance monitoring. Clinical Trial: Not applicable (observational study; no intervention).