Retinal artificial intelligence as a gatekeeper for Coronary Artery Calcium Score (CACS) referral: a novel strategy for identifying clinically relevant coronary atherosclerosis

Background/Introduction Prediction of clinically relevant coronary atherosclerosis based solely on conventional cardiovascular (CV) risk factors remains challenging. While CACS is effective for risk stratification, its routine use as a primary screen is limited by radiation exposure, cost, and availability. Conversely, relying exclusively on CV risk factors as a gatekeeper for CACS may misclassify a substantial number of patients with relevant coronary disease. Purpose To evaluate whether an artificial intelligence (AI)-based retinal cardiovascular score (Retinal AI score) provides incremental value and can serve as an effective hybrid gatekeeper to optimise CACS referral. Methods This study analyzed a cohort of 1,205 individuals who underwent Coronary Computed Tomography Angiography (CCTA) and same day retinal imaging at a high-volume medical imaging center in Greece. Clinically relevant coronary atherosclerosis was defined as a Coronary Artery Disease Reporting and Data System (CAD-RADS) score ≥2 or the presence of high-risk plaque. Incremental predictive value was assessed comparing the area under the curve (AUC) across nested models: 1) CACS alone; 2) CACS + Clinical Rule; 3) CACS + Clinical Rule + Retinal AI score. Three CACS referral strategies were compared: 1) Universal CACS strategy (all patients undergo CACS); 2) Clinical Rule strategy (referral based on diabetes or ≥2 other CV risk factors [hypertension, dyslipidaemia, smoking]); and 3) Hybrid strategy (referral based on the Clinical Rule or a high Retinal AI score [upper two tertiles]). Performance (sensitivity, false omission rate [FOR], accuracy) was assessed via bootstrap resampling (2,000 iterations) against the CCTA reference. Results Among 1,205 participants (54.5±9.8 years, 65.8% male), 25.3% had clinically relevant coronary atherosclerosis. In the stepwise incremental analysis, the AUC significantly improved from 0.757 for CACS alone to 0.777 with the Clinical Rule and further increased to 0.789 upon incorporation of the Retinal AI score (ΔAUC = +0.0118 vs. CACS + Clinical Rule model, p=0.005). Regarding CACS referral strategies, compared to the Universal CACS strategy, the Clinical Rule strategy reduced CACS workload by 42.9% but compromised diagnostic performance, with sensitivity dropping to 48.5% (vs. 62.0%) and the FOR increasing to 15.8% (vs. 12.6%). In contrast, the Hybrid strategy reduced workload by 18.3% while maintaining a sensitivity of 59.3%, an FOR of 13.2%, and an overall accuracy of 82.41%, which was not statistically different from that of the Universal CACS strategy (82.49%; p=0.920). Conclusion(s) Using conventional CV risk factors alone to guide CACS referral misses many patients with relevant coronary atherosclerosis. Integrating a highly accessible AI-based retinal CV score as a hybrid gatekeeper optimises CACS referral—achieving an 18.3% reduction in unnecessary testing while preserving diagnostic accuracy.