P123 - PREDICTION OF ATRIAL FIBRILLATION FOLLOWING CARDIAC SURGERY USING DEEP LEARNING, CLINICAL MODELS, AND POLYGENIC SCORES

Background: Post-operative AF (POAF) is the most common complication following cardiac surgery. It is associated with prolonged hospital stay and increased in-hospital mortality. Identifying individuals at risk of POAF is an unmet clinical need. This study evaluates the performance of multiple modalities in predicting POAF, namely, CHARGE-AF clinical score, genome-wide AF polygenic score (PGS-AF), and electrocardiogram-based deep learning (ECG-AI).

METHODS AND RESULTS: In a retrospective study, over 20,000 participants from the Montreal Heart Institute (MHI) Biobank were screened. Individuals with available genotypic data were included in the study if they had undergone cardiac surgery at MHI between 2005 and 2020 and had no documented AF diagnosis prior to surgery. Patients who underwent transcatheter interventions and cardiac transplantations were excluded. POAF was defined as new-onset AF or atrial flutter first documented on an ECG or in medical records within 30 days following cardiac surgery. CHARGE-AF score components were retrieved from medical records closest to the surgery date. A genome-wide PGS-AF (PGS catalog ID PGS000016) was computed for included patients. An ECG-AI model, previously developed to predict 5-year incident AF by detecting subtle features on 12-lead ECG (Jabbour et al, Eur Heart J 2024), was used to derive a pre-operative probability of POAF. A total of 2,340 patients were included. Median age was 66 years at the time of surgery, 80% were male, and 99% of European ancestry. Most surgeries were coronary artery bypass grafts, either in isolation (53%) or with valve intervention (14%); the remainder were mainly isolated valve interventions (30%). The primary outcome, POAF, was observed in 871 (37%) patients. POAF risk was significantly higher among patients with PGS-AF above the 95th and 99th percentiles, corresponding to 2.3-fold and 3.7-fold increases, respectively (P < 0.001). Compared to CHARGE-AF only, adding PGS-AF resulted in a significant improvement of POAF prediction (C-statistic 0.68 vs. 0.65; likelihood ratio P< 0.001). In a subgroup analysis of 1,855 patients with available ECG acquired at MHI prior to surgery, the ECG-AI model identified 36% of patients at high risk of incident AF. Addition of ECG-AI high-risk stratification on top of both CHARGE-AF and PGS-AF was independently associated with increased odds of POAF (OR 1.85, 95% CI 1.50–2.28, P< 0.001).

Conclusion: PGS-AF improved prediction of POAF beyond CHARGE-AF, and addition of ECG-AI high-risk stratification further enhanced model performance. Combining deep learning-based ECG analysis with genetic and clinical risk factors may refine POAF risk assessment and enable targeted perioperative interventions.