Opportunistic Screening of Coronary Artery Calcium on Non-Gated Routine Chest CT Using Artificial Intelligence: Retrospective External Validation and Clinical Risk Stratification

The study analyzes a retrospective cohort of routine clinical CT examinations at University Hospital Cologne. Data originate from the hospital information system and Picture Archiving and Communication System (PACS) and are provided in pseudonymized form via the Medical Data Integration Center (MeDIC), acting as an independent trusted third party; the re-identification key remains under the sole control of MeDIC. Deep-learning inference is performed locally on isolated, access-controlled graphics processing unit (GPU) clusters of the institution (privacy by design / zero data retention); an open-weight model (Swin-UNETR) is used.

Three research questions are addressed in three analytic cohorts:

• Validation (n ≈ 150): diagnostic agreement of the automatically extracted AI-CAC score (from the non-gated CT) with the reference Agatston score from a paired ECG-gated cardiac CT acquired within ≤ 12 months.
• Dialysis (n ≈ 300): 150 hemodialysis patients plus 150 matched kidney-healthy controls with serial non-gated CTs; annualized calcification progression rate and all-cause mortality.
• Emergency department (n ≈ 1,500): patients > 50 years with non-gated chest CTs from the Emergency Department (without a primary cardiac focus); occurrence of in-hospital major adverse cardiac event (MACE) or cardiovascular readmission within 12 months.

Extracted data include demographics (age at examination, sex), cardiovascular risk factors and comorbidities (ICD-10), long-term medication, laboratory values, examination metadata (date, scanner manufacturer, kilovolt peak (kVp), slice thickness), and outcome data (mortality, cardiovascular events, readmissions). Statistical analysis uses Spearman correlation, Cohen's kappa and Bland-Altman analysis for method comparison; t-test / Mann-Whitney-U for group differences in progression; and Kaplan-Meier (log-rank) plus multivariable Cox proportional-hazards and logistic regression for outcome prediction. Legal basis: § 6 (1) no. 2 Health Data Use Act of Germany (GDNG) in conjunction with Art. 9 (2) (j) and Art. 89 (1) GDPR (research privilege); no individual consent (disproportionate effort, Art. 14 (5) (b) GDPR). The AI (artificial intelligence) model carries no CE-marking and is used strictly as a research tool; AI-CAC scores are not systematically fed back into clinical care.