Association between a vision-capable large language model’s image-derived score and CT-defined coronary artery calcium category on routine chest radiographs: a single-center retrospective study

Abstract Background Coronary artery calcium (CAC) is an established marker of cardiovascular risk, but CT-based CAC scoring is not universally performed. This study evaluated whether a vision-capable large language model (LLM) can classify clinically significant CAC (CAC > 100) from routine posteroanterior chest radiographs, using CT-derived Agatston score as the reference standard. Methods This retrospective single-center diagnostic accuracy study included 349 patients who underwent both non-contrast cardiac CT for CAC scoring and chest radiography within 6 months. Patients were classified as CAC 0–100 or CAC > 100. Preprocessed chest radiographs were analyzed by GPT-5.5 in a zero-shot setting using a structured prompt. The model estimated the score of CAC > 100, assigned a binary class, and reported visible calcification findings. Diagnostic performance was assessed against CT-derived Agatston classification. Results Of 349 patients, 131 had CAC > 100. Using its native binary output, the LLM correctly classified 292 patients (accuracy 83.7%; sensitivity, specificity, positive predictive value, and negative predictive value 72.5%, 90.4%, 81.9%, and 84.5%, respectively). The threshold-independent area under the receiver operating characteristic curve was 0.771. Applying the continuous model-generated score at a prespecified 0.50 threshold yielded 36.6% sensitivity and 94.0% specificity. Model-generated scores were significantly higher in patients with CAC > 100 but were poorly calibrated (calibration slope 2.49), with substantial overlap in the intermediate range. Conclusion A vision-capable LLM produced image-derived scores that were associated with CT-defined CAC category in this selected cohort, without task-specific training. This exploratory, hypothesis-generating finding does not establish clinical feasibility for opportunistic risk flagging, which would require evaluation in an unselected radiography population, human-reader benchmarking, and prospective validation before any clinical use.