Machine Learning-Based CAC Score Prediction from Clinical Variables as a Prognostic Indicator

This retrospective cohort study developed and validated a machine learning model using nine clinical variables (including age, sex, BMI, blood pressure, lipids, glucose, and smoking status) to predict severe coronary artery calcium (CAC) score in 5,785 patients referred for invasive coronary angiography, evaluating its prognostic value for adverse cardiovascular events.

The ML-CAC model achieved an AUC of 0.753 for predicting severe CAC, and higher ML-CAC scores were significantly associated with increased risk of adverse cardiovascular events at follow-up. The model used clinical variables available at routine assessment rather than actual CT-derived calcium scores.

The motivation for this approach: actual CAC scoring requires CT imaging and additional radiation exposure. A machine learning model that estimates CAC risk from clinical variables could theoretically triage which patients most need actual CAC measurement — directing the imaging cost and radiation to those most likely to have elevated scores.

The AUC of 0.753 is moderate — it indicates the model can rank patients by calcium burden better than chance but misclassifies a substantial proportion. The clinical utility depends on whether this discriminative performance is sufficient to change management: would a high ML-CAC score send a patient to CT scanning who otherwise wouldn’t go? Given that CAC scoring is inexpensive (~$100), low-radiation, and highly informative for clinical decision-making in intermediate-risk patients, the incremental value of a clinical prediction model with moderate AUC is uncertain.

We rate the evidence limited. A retrospective ML study developing a clinical variable-based CAC score predictor with moderate accuracy — methodologically interesting but unclear whether the prediction model provides meaningful value over direct CAC testing.