Development and validation of machine learning-based risk prediction models of coronary artery disease using Porphyromonas gingivalis concentration

Porphyromonas gingivalis (P. gingivalis) is a highly prevalent pathogen in dysbiotic dental biofilms in periodontitis. The present study aimed to investigate the independent predictive value of P. gingivalis concentration for coronary artery disease (CAD) and develop a machine learning (ML)-based risk prediction model incorporating this biomarker. This pilot study enrolled 36 participants undergoing diagnostic coronary angiography, comprising 18 CAD patients (≥ 50% stenosis) and 18 controls (no significant stenosis). Circulating P. gingivalis DNA was quantified using qPCR. Traditional CAD risk factors (age, sex, BMI, smoking, alcohol consumption, lipid profiles, glucose, inflammatory markers) were incorporated as candidate predictors. Three machine learning algorithms-Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM)-were developed. Feature selection was performed using the Boruta algorithm. The added predictive value of P. gingivalis concentration was quantified using Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) by comparing models with and without this biomarker. Model performance was evaluated using Area Under Curve (AUC), Brier Score, calibration curve, Hosmer-Lemeshow test and Decision Curve Analysis (DCA). Internal validation was performed using 1000 bootstrap repetitions. The final model was visualized as an interactive nomogram. The RF model exhibited superior performance (AUC = 0.965, Brier score = 0.102), excellent calibration, and significantly outperforming the XGBoost model (AUC = 0.859, Brier score = 0.169) and the SVM model (AUC = 0.148, Brier score = 0.302). SHAP analysis confirmed that P. gingivalis concentration was a primary contributor to the model’s predictions. The NRI and IDI for P. gingivalis concentration were 0.444 (95%CI: 0.134–0.755, P < 0.05) and 0.444 (95%CI: 0.273–0.615, P < 0.05), respectively, underscoring its significant predictive contribution to CAD risk assessment. A robust CAD risk prediction model incorporating P. gingivalis concentration was successfully developed using an RF algorithm. Systematic evaluation confirmed the model’s high clinical utility and established the potential independent predictive value of P. gingivalis concentration for CAD risk prediction.