Machine Learning for the Prevalence and Severity of Coronary Artery Calcification in Nondialysis Chronic Kidney Disease Patients: A Chinese Large Cohort Study

This study sought to determine whether machine learning (ML) can be used to better identify the risk factors and establish the prediction models for the prevalence and severity of coronary artery calcification (CAC) in nondialysis chronic kidney disease (CKD) patients and compare the performance of distinctive ML models with conventional logistic regression (LR) model. MATERIALS AND METHODS: In all, 3701 Chinese nondialysis CKD patients undergoing noncontrast cardiac computed tomography (CT) scanning were enrolled from November 2013 to December 2017. CAC score derived from the cardiac CT was calculated with the calcium scoring software and was used to assess and stratify the prevalence and severity of CAC. Four ML models (LR, random forest, support vector machine, and k-nearest neighbor) and the corresponding feature ranks were conducted. The model that incorporated the independent predictors was shown as the receiver-operating characteristic (ROC) curve. Area under the curve (AUC) was used to present the prediction value. ML model performance was compared with the traditional LR model using pairwise comparisons of AUCs. RESULTS: Of the 3701 patients, 943 (25.5%) patients had CAC. Of the 943 patients with CAC, 764 patients (20.6%) and 179 patients (4.8%) had an Agatston CAC score of 1 to 300 and ≥300, respectively. The primary cohort and the independent validation cohort comprised 2957 patients and 744 patients, respectively. For the prevalence of CAC, the AUCs of ML models were from 0.78 to 0.82 in the training data set and the internal validation cohort. For the severity of CAC, the AUCs of the 4 ML models were from 0.67 to 0.70 in the training data set and from 0.53 to 0.70 in the internal validation cohort. For the prevalence of CAC, the AUC was 0.80 (95% confidence interval [CI]: 0.77-0.83) for ML (LR) versus 0.80 (95% CI: 0.77-0.83) for the traditional LR model (P=0.2533). For the severity of CAC, the AUC was 0.70 (95% CI: 0.63-0.77) for ML (LR) versus 0.70 (95% CI: 0.63-0.77) for traditional LR model (P=0.982). CONCLUSIONS: This study constructed prediction models for the presence and severity of CAC based on Agatston scores derived from noncontrast cardiac CT scanning in nondialysis CKD patients using ML, and showed ML LR had the best performance.