Improving Cardiovascular Disease Prediction Using Automated Coronary Artery Calcium Scoring from Existing Chest CTs.

Area of Science:

• Cardiology
• Medical Imaging
• Machine Learning

Background:

• Cardiovascular disease (CVD) prediction models are integral to clinical practice and guidelines.
• Coronary artery calcium (CAC) is a validated marker for coronary atherosclerotic disease.
• Current guidelines often exclude CAC from prediction models due to costs and insufficient evidence of improvement.

Purpose of the Study:

• To evaluate if automatically extracted CAC scores from existing chest CT scans enhance CVD prediction models.
• To compare the performance of the standard American Heart Association/American College of Cardiology (AHA/ACC) 2013 pooled cohort equations (PCE) with an augmented model incorporating CT-derived CAC scores.

Main Methods:

• A retrospective cohort study of 14,135 patients aged 40-79 with pre-2012 chest CT scans.
• CAC scores were automatically extracted using machine learning from existing chest CTs.
• Compared the predictive performance of the standard PCE model versus an augmented-PCE model (PCE + CAC) for 5-year CVD events (until 2017).

Main Results:

• The augmented-PCE model demonstrated significant improvements in c-statistic (0.64 to 0.69), sensitivity (53% to 57%), and specificity (67% to 70%).
• Positive predictive value increased from 5% to 6%, and negative predictive value from 97.7% to 97.9%.
• The categorical net reclassification index showed a 7.4% improvement, indicating better risk prediction.

Conclusions:

• Automatically generated CAC scores from readily available CT scans can significantly improve CVD risk prediction.
• This approach circumvents the costs and radiation exposure associated with dedicated CAC scans.
• Integrating CT-derived CAC scores offers a net gain in predictive accuracy for cardiovascular events.