Machine Learning and Explainable Artificial Intelligence for Predicting 10-Year Mortality after Cardiac Surgery: A Retrospective Cohort Study

Abstract Background Preoperative neutrophil-to-lymphocyte ratio (NLR) is associated with short-term outcomes after cardiac surgery, but its value for long-term mortality prediction remains unclear. We developed a machine learning model incorporating NLR to predict 10-year all-cause mortality after cardiac surgery, and used explainable artificial intelligence to interpret the role of NLR. Methods We included adult patients undergoing elective cardiac surgery from the MIMIC-IV database. The primary outcome was 10-year all-cause mortality. Six machine learning models were compared using grid search, stratified 5-fold cross-validation, and bootstrap internal validation. The best model was interpreted using SHAP on the test set. The NLR cutoff was identified by piecewise regression on the training set. Calibration curves and decision curve analysis were used to evaluate model performance. Results Among 4,891 patients (median age 68 years), 222 (4.54%) died within 10 years. After tuning, XGBoost performed best (optimism-corrected AUC = 0.815; 95% CI: 0.758–0.861), outperforming LASSO regression (AUC = 0.780). NLR alone had an AUC of 0.709, similar to APACHE III (0.732; P = 0.630). Combining NLR with APACHE III improved the AUC to 0.761 (P = 0.002). SHAP analysis identified APACHE III (0.371) and NLR (0.291) as the top two predictors. Piecewise regression identified an NLR cutoff of 6.5. At a 15% risk threshold, the positive predictive value was 23.3% (5.1-fold enrichment). Decision curve analysis showed net benefit across thresholds of 5%-30%. Conclusions Preoperative NLR is an independent predictor of 10-year mortality after cardiac surgery, with performance comparable to APACHE III. The tuned XGBoost model showed good overall performance. An NLR threshold of 6.5 was identified and supported by SHAP analysis. NLR may complement existing risk scores for long-term risk stratification, but should not be used as a standalone tool.