Machine Learning Models for Identifying the Risk of Chronic Kidney Disease in Patients with Coronary Heart Disease: A Retrospective Study

Purpose: Coronary heart disease (CHD) has a significant co-morbid association with chronic kidney disease (CKD), but identification tools for the risk of concomitant CKD in patients with CHD are still lacking. The purpose of this research was to construct machine learning (ML) models for identifying undetected CKD in CHD patients. Methods: 1786 CHD patients undergoing coronary intervention were retrospectively included. Lasso regression and multifactor logistic regression were used to screen feature variables. Five ML models, ie, logistic regression (LR), support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), and extreme gradient boosting (XGBoost), were constructed. Participants were divided into the training set and validation set in a 7:3 ratio. The evaluation metrics included the area under the curve, calibration curve, and decision curve. Results: Totally, 1786 CHD patients were enrolled and split into training (70%) and validation (30%) sets. The prevalence of CKD was 21.8% (390/1786). Multivariate logistic regression analysis showed that men, advanced age, hypertension, diabetes mellitus, history of atrial fibrillation (AF), high Gensini, low hemoglobin, low plateletcrit (PCT), high triglycerides (TG), high lipoprotein(a) (Lp(a)), hyperkalemia, high uric acid to albumin ratio (UAR), high systemic inflammation response index (SIRI), low lymphocyte to monocyte ratio (LMR), and high apolipoprotein B to apolipoprotein A1 (ApoB/ApoA1) ratio were the key clinical and laboratory test indicators of CKD. The XGBoost model performed optimally in the validation set (AUC=0.909, 95% CI 0.881 -0.937). SHapley Additive explanation analysis identified UAR, hypertension, Gensini score, age, and SIRI as the top 5 key features. Conclusion: The XGBoost model constructed on routine clinical data was effective in identifying CKD risk in CHD patients, with UAR as a novel strong predictor. Decision curve analysis confirmed the clinical utility of the model, indicating that it may be used to guide decisions for enhanced monitoring and early intervention over a wide range of risk thresholds.