Interpretable machine learning-derived risk score for significant coronary artery calcification in a real-world clinical cohort
J Jang, J S Park, J Chung, K H Kim, E M Lee, M Y RheeAbstract
Background/Introduction
Coronary artery calcium score (CACS) ≥100 is a validated marker of subclinical coronary atherosclerosis and is associated with increased risk of future cardiovascular events. However, assessment requires non-contrast cardiac computed tomography, which is not universally available.
Purpose
To develop and validate an interpretable clinical risk score for predicting significant coronary artery calcification using routinely collected clinical data.
Methods
A real-world cohort of 4,399 individuals with linked laboratory, biometric and computed tomography data was analysed, of whom 2,801 had available CACS measurements. The primary outcome was CACS ≥100. Feature selection was guided by Shapley Additive Explanations to enhance interpretability. A multivariable logistic regression model was constructed and internally validated. Regression coefficients were scaled and rounded to derive a simplified point-based clinical risk score. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC).
Results
Age, inflammatory markers (C-reactive protein and neutrophil-to-lymphocyte ratio), lipid parameters, blood pressure indices, body mass index and estimated glomerular filtration rate were independently associated with CACS ≥100. The logistic regression model demonstrated moderate discrimination (AUC 0.779). The simplified point-based risk score retained acceptable predictive performance (AUC 0.720). Although a gradient boosting model achieved higher discrimination (AUC 0.941), the interpretable risk score was prioritised to facilitate clinical applicability and transparent risk estimation. Risk stratification using the score demonstrated a graded increase in the prevalence of CACS ≥100 across ascending risk categories.
Conclusions
An interpretable risk score derived from routine clinical data enables identification of individuals at increased likelihood of significant coronary artery calcification. This pragmatic approach may support cardiovascular risk stratification in settings where cardiac computed tomography is not readily accessible.CACS_ROC_comparison_ESC2026.jpg