Interpretable machine learning for postoperative atrial fibrillation after noncardiac surgery

Postoperative atrial fibrillation (POAF) is an important complication after non-cardiac surgery and is associated with adverse short- and long-term outcomes. Accurate perioperative risk stratification remains challenging because POAF risk reflects both baseline atrial substrate and dynamic perioperative physiologic stress. We conducted a retrospective cohort study using the INSPIRE perioperative database to develop interpretable prediction models for clinically documented postoperative atrial fibrillation (POAF) after non-cardiac surgery. POAF was defined as new-onset AF captured during the index hospitalization by electrocardiography or rhythm-monitoring documentation; because a uniform cohort-wide measure of monitoring intensity was not available, the endpoint was interpreted as clinically detected POAF. Candidate predictors were grouped by acquisition window into preoperative, intraoperative, and early postoperative (≤ 24 h) variables to support dynamic perioperative risk assessment rather than a purely preoperative screen. Feature selection was performed with Boruta, model explanation with SHapley Additive exPlanations (SHAP), and classifier comparison with LASSO logistic regression, AdaBoost, LightGBM, XGBoost, and CatBoost. Because POAF was uncommon, performance was evaluated with AUROC, sensitivity, specificity, PPV, NPV, and F1 score after stratified splitting and threshold evaluation; synthetic resampling was not used in the primary analysis. Among 99,887 patients, 1,058 (1.1%) developed clinically documented POAF. Compared with controls, POAF cases were older (median 70 vs. 55 years) and showed substantially greater burdens of hypotension and hypoxemia, including more frequent and prolonged MAP < 65 mmHg and SpO₂ <90%. Boruta retained average diastolic blood pressure, average heart rate, and left atrial enlargement as informative features. SHAP identified valvular disease, surgical duration, cardiopulmonary instability, and related early postoperative events as major contributors, consistent with a dynamic perioperative risk model. Discrimination was good across classifiers, with AUCs ranging from 0.80 to 0.85. AdaBoost and CatBoost achieved the best overall performance (both AUC 0.85), while CatBoost showed higher specificity and PPV. In a large non-cardiac surgical cohort, clinically documented POAF was associated with both baseline atrial substrate and perioperative cardiopulmonary stress. Interpretable ML models, particularly boosting-based methods, showed good discrimination and may support dynamic perioperative risk updating and targeted postoperative surveillance, although monitoring heterogeneity, class imbalance, calibration, and external validation remain important considerations.