Mortality prediction in the elderly patients with coronary artery disease and atrial fibrillation: a retrospective machine learning approach

Abstract

Background Elderly patients with coexisting coronary artery disease (CAD) and atrial fibrillation (AF) are at significantly increased risk of mortality. Accurate risk stratification is crucial for improving clinical management, yet a dedicated predictive tool for this specific population is lacking. The widely used CHA2DS2-VASc score demonstrates limited performance in predicting all-cause mortality in this complex comorbid group. Methods A cohort of elderly inpatients (≥65 years) diagnosed with CAD and AF were retrospectively enrolled from the Department of Cardiology at the Chinese PLA General Hospital between January 2010 and December 2017. Baseline clinical data during hospitalization were collected, and all patients were followed up for all-cause mortality. Patients were randomly divided into a training set (70%) and a validation set (30%). Variable selection was performed using LASSO-Cox regression. Predictive models were established through a Cox proportional hazards model (via LASSO-Cox) to predict all-cause mortality risk. To assess model robustness and address potential overfitting, we performed 10-fold cross-validation and bootstrap resampling (500 replicates) with optimism-corrected C-statistics. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curves, and decision curve analysis (DCA), and compared against the CHA2DS2-VASc score. Results A total of 1678 elderly patients with CAD and AF were randomly divided into a training set (n = 1174) and a validation set (n = 504). Through LASSO-Cox regression, 17 variables were identified as predictors associated with all-cause mortality. Two distinct models were developed: LASSO-Cox Model A included factors such as age, left ventricular ejection fraction, blood glucose, plasma fibrinogen, D-dimer, prothrombin time, NT-proBNP, hemoglobin, and hematocrit; LASSO-Cox Model B incorporated all variables from Model A plus acute myocardial infarction, history of myocardial infarction, renal insufficiency, heart failure, diabetes, stroke, chronic obstructive pulmonary disease, and malignant tumor. The analysis of Cox proportional hazards regression models showed that in the training sets, the Area Under Curves (AUC) of model A, model B and CHA2DS2-VASc scores for predicting 1-year all-cause death were 0.83, 0.85 and 0.66, respectively; the AUC for 5-year all-cause death were 0.74, 0.76 and 0.62, respectively. In the validation sets, the AUC of model A, model B and the CHA2DS2-VASc score for predicting 1-year all-cause mortality were 0.79, 0.78 and 0.57, respectively; the AUC for 5-year all-cause death were 0.74, 0.75 and 0.57, respectively. The 10-fold cross-validation yielded a mean AUC of 0.83 (SD: 0.04) for 1-year mortality, and bootstrap resampling produced an optimism-corrected C-statistic of 0.679 (95% CI: 0.651–0.708), confirming that the model maintains clinically meaningful discriminative ability after accounting for optimism. Model B demonstrated better calibration and provided greater net clinical benefit in DCA. Conclusion The LASSO-Cox machine learning model demonstrates superior predictive performance for all cause mortality relative to the traditional CHA2DS2-VASc score in elderly patients with CAD and AF.