Fully automatic coronary artery calcium score using CT and artificial intelligence: external validation on a danish population

The Coronary Artery Calcium Score (CACS) is a well validated CT-derived quantitative marker for Major Adverse Cardiac Events (MACE), providing incremental value beyond traditional risk factors.[1-3] Current clinical practice is, however, still semi-automatic making the process time-consuming, and only performed in CT scans with the purpose of calculating CACS. Using Artificial Intelligence (AI), CACS could be calculated automatically. Recent AI models are showing promising results, often getting similar or even improved CACS calculation compared to humans [1-3], however there is a large diversity across the literature, meaning high variation in study population and study design, which might result in models not generalizing well to other patient populations or CT scanner protocols [3]. The AI-CAC model developed by Hagopian et. Al. in 2025 [1], trained on a U.S. cohort to calculate CACS on non-gated, non-contrast CT scans, displayed a high level of agreement with manually reported CACS on gated CT scans (Cohen’s κ = 0.71) tested on 794 patients. The AI-CAC model is openly published, making external validation possible. The purpose of this study was to validate the performance of the pretrained AI-CAC model [1] on an external dataset. The AI-CAC model was applied on an external dataset of 2474 patients CT scans with CACS reported by human experts. The dataset consisted of a combination of patients from the Danish trials; DanNicad 1, 2, and 3, with patients being enrolled with the suspicion of coronary artery disease. CT scans were non-contrast, full field of view, and with slice thickness between 2–4mm. The CACS were divided into four groups of 0, 1-100, 101-400, and >400. Linearly-weighted Cohen’s κ: 0.75 (95% CI 0.73–0.77) and R=0.84 shows strong agreement between the predictions of AI-CAC and the manually reported CACS, as seen on figure 1 and 2. The performance of AI-CAC on the external DanNicad cohort even seems to have increased agreement compared to the original paper [1], which might be due to CT scans being ECG-gated in the DanNICAD trials. The sensitivity and specificity of predicting CACS >400 were 0.90 and 0.94 respectively. Computational time took an average of 63 seconds per patient, run on a desktop PC with an RTX4080 GPU. The model was able to automatically calculate the CACS on the external dataset and agreement coefficient κ = 0.75 was achieved between the AI prediction and human expert reported CACS. This demonstrates AI-CAC’s ability to be generalizable and in agreement with human experts, which could help push the clinical use of AI for automatic CACS. A limitation of the AI-CAC model is that the model was trained using only full field of view. Cardiac scans are often focused on the cardiac region. This reduced available CT data from 4407 patients to 2474 for this validation, as many studies only had a cardiac focused series for CACS. Future work should address this by training on mixed views and zooms.