AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
A Deep Learning Risk Score Integrating Polygenic Risk and Longitudinal EHR History Improves CAD Risk Stratification Beyond PREVENT
By Yang L, Ratman D, Lavertu A et al.
Introduction: The AHA PREVENT equations use traditional clinical risk factors to assess coronary artery disease (CAD) risk and guide statin therapy. However, the model does not take into account the risk conferred by CAD Polygenic Risk Score (PRS) or information from full longitudinal electronic health record (EHR) data. It also requires complete clinical inputs, which limit its applicability in EHR systems where missing data is common. We tested whether integrating CAD PRS with a deep learning model pre-trained on EHR history improves CAD risk prediction among statin naive, PREVENT-eligible individuals and further evaluated the model among individuals who would not be eligible for risk assessment with PREVENT due to missing input data or values beyond allowed boundaries. Methods: We developed a deep learning Cox model pre-trained on a large corporate EHR corpus (193,273 CAD cases; 1,067,458 controls) and fine-tuned it in All of Us participants with whole-genome sequencing and EHR data. Baseline was January 1, 2019; a 2-year clinical lookback was used; and participants on lipid-lowering therapy or with CAD diagnoses prior to baseline were excluded. The iPRS_EHR model encodes the full temporal EHR with genome-wide CAD PRS and is robust to missing clinical inputs. Data were split into training (n=49,239), validation (n=37,101), and test (n=37,014) sets. In the PREVENT-eligible test subset (n=4,891; 332 events) we compared the iPRS_EHR to PREVENT and the linear, additive model combining PREVENT with CAD PRS (iPRS_PREVENT). To enable fair comparison of absolute risk estimates, the PREVENT model (10-year) was recalibrated using the training set. Discrimination was evaluated by C-index and Net Reclassification Improvement (NRI) at clinically recommended intermediate (5%) and high risk (10%) thresholds. Results: Among 32,123 test-set participants without computable PREVENT scores, including 1,676 incident CAD cases, the iPRS_EHR model achieved the C-index of 0.785 (CI: 0.774-0.794), demonstrating strong discrimination when clinical inputs were incomplete. In the PREVENT-eligible subset, the C-index improved stepwise from PREVENT to iPRS_PREVENT and iPRS_EHR (0.709, 0.714, and 0.748, respectively). Compared with PREVENT, iPRS_PREVENT achieved NRIs of +0.4% and +2.3% at the 5% and 10% thresholds, while the iPRS_EHR model achieved +1.1% and +8.4%. Extrapolating from U.S. incidence and mortality, this improvement may correspond to ~23,000 cardiovascular deaths prevented over 5 years. Conclusions: PRS integration improves PREVENT-based CAD risk prediction, and our pre-trained deep learning model combining genome-wide PRS with longitudinal EHR history provides further improvement beyond iPRS_PREVENT. Its discrimination in participants without computable PREVENT scores supports broader applicability and scalable precision cardiovascular risk stratification.