To develop and validate machine learning models for predicting 30-day postoperative pulmonary complications (PPCs) in patients aged ≥65 years undergoing abdominal surgery.
Approach:
Study Design: Retrospective cohort study including 2,456 patients in a development/internal cohort and 542 patients in an independent external-validation cohort.
Model Comparison: Six algorithms were compared: logistic regression, random forest, support vector machine, neural network, XGBoost, and LightGBM.
Performance Evaluation: Model performance was assessed using discrimination, calibration, decision curve analysis, and external validation, with SHAP analysis for interpretability.
Key Findings:
PPCs occurred in 17.3% of the development cohort and 19.4% of the external-validation cohort.
XGBoost demonstrated the best performance with AUCs of 0.856 and 0.821 in the independent test and external-validation cohorts, respectively.
At a 20% risk threshold, sensitivity was 87.5%, specificity was 60.9%, positive predictive value (PPV) was 32.0%, and negative predictive value (NPV) was 95.9%.
Key contributors to PPC risk included ASA physical status, COPD, upper abdominal surgery, age, emergency surgery, albumin, and surgical duration.
Interpretation:
The model effectively stratified patients into low-, moderate-, and high-risk groups with observed PPC rates of 6.9%, 24.4%, and 37.6%, respectively.
Limitations:
The study is retrospective and may not capture all relevant clinical variables.
External validation was limited to one additional cohort.
Conclusion:
An interpretable gradient-boosting model may support risk-stratified perioperative assessment for elderly patients undergoing abdominal surgery, pending prospective multicenter validation.
