The Predictive Value of the Pan-Immune-Inflammation Value for Atrial Fibrillation Risk in Patients with Coronary Artery Disease: A Multicenter Machine Learning Study

Background: Atrial fibrillation (AF) is a common arrhythmia among patients with coronary heart disease (CHD), and inflammatory response plays a key role in its pathogenesis. The pan-immune-inflammation value (PIV), a novel composite marker reflecting systemic inflammation, has not been fully investigated for its predictive value in AF among CHD patients. Methods: This multicenter retrospective study enrolled patients diagnosed with CHD by coronary angiography from two tertiary hospitals. Participants were categorized into AF and non-AF groups. Clinical characteristics and laboratory data were collected. Feature selection was performed using multivariate logistic regression, and significant predictors were incorporated into two models: extreme gradient boosting (XGBoost) and multilayer perceptron (MLP). Model performance was evaluated by area under the ROC curve (AUC) and calibration analysis. Model interpretability was assessed using SHAP (SHapley Additive exPlanations) values, and partial dependence plots (PDPs) were applied to explore variable interactions. Results: Compared with the non-AF group, the AF group had significantly higher levels of PIV, age, AST, WBC, and TBIL. Logistic regression identified PIV, age, and diabetes as independent predictors of AF, while sex, left main coronary artery disease (LM), and AST showed borderline significance. The XGBoost model achieved superior performance (AUC = 0.79 in training and 0.73 in testing) compared to the MLP model (AUC = 0.75 and 0.69, respectively), with better calibration consistency. SHAP analysis indicated that PIV was the most influential feature, with higher values associated with increased AF risk. PDPs further demonstrated synergistic effects between PIV and other key variables. Conclusion: PIV is a valuable predictor of AF in CHD patients. The XGBoost model outperformed the deep learning model in this context and may serve as a robust tool for individualized AF risk assessment.