Multimodal prediction of atrial fibrillation following cardiac surgery using clinical and polygenic scores
G Jabbour, A Jeuken, P Jorda, A M Petzl, S J Jurgens, D Bouchard, Y Lamarche, R Avram, L Rivard, P Khairy, P E Noly, L Macle, J C Tardif, J Cadrin-Tourigny, R TadrosAbstract
Background/Introduction
Postoperative atrial fibrillation (POAF) is the most common complication following cardiac surgery, prolongs hospital stay and increases risk of stroke and death.
Purpose
To develop a prediction model for POAF integrating clinical variables and polygenic risk, and to compare its performance with existing atrial fibrillation (AF) prediction models, including CHARGE-AF, polygenic risk scores (PGS-AF), and an ECG-based artificial intelligence (ECG-AI) model.
Methods
This was a prospective hospital biobank-based cohort study with retrospective adjudication of POAF. We included 2,340 participants in a Biobank at our institution who underwent cardiac surgery between 2005 and 2020 and did not have AF prior to surgery. POAF was defined as new-onset AF occurring within 30 days after surgery. PGS-AF was calculated using the Multi-t-EUR model, a European ancestry-specific tuned score integrating individual polygenic scores for AF and related traits (1). A least absolute shrinkage and selection operator (LASSO) logistic regression model and a simplified point-based score (PAC-AF), incorporating PGS-AF percentiles, age ≥65 years, and surgical characteristics, were derived using clinical variables and PGS-AF (Figure, panel B). Model tuning was conducted using the training set (70%). Performance assessment is reported on the test set (30%). The ECG-AI model used for comparison was previously developed to predict long-term incident AF rather than POAF (2). Performance was assessed using odds ratios (OR) between risk categories, calibration plots and the area under the receiver operating characteristic curve (AUC).
Results
Among 2,340 patients, 871 (37.2%) developed POAF (Table). In the test cohort, the LASSO model demonstrated the highest discrimination (AUC 0.72, 95% CI 0.68–0.75), with PGS-AF emerging as the most powerful independent predictor (OR 1.83, 95% CI 1.64–2.08 per 1-SD increase). The simplified point-based PAC-AF score yielded an AUC of 0.68 (95% CI 0.64–0.72), with no statistically significant difference compared to LASSO model (DeLong p=0.19). POAF prevalence increased stepwise across PAC-AF score categories (Figure), from 21% in patients with ≤2 points to 83% in those with ≥6 points, corresponding to an odds ratio of 18.5 (95% CI 5.2–66.2). CHARGE-AF (AUC 0.64, 95% CI 0.61–0.66), PGS-AF (AUC 0.64, 95% CI 0.62–0.67), and ECG-AI (AUC 0.65, 95% CI 0.62–0.67) showed lower discrimination. The LASSO and PAC-AF models showed good calibration (slopes 1.14 and 1.08, respectively).
Conclusion
A multimodal model integrating clinical variables and polygenic risk demonstrated good performance in predicting POAF, with PGS-AF emerging as the most powerful independent predictor. The PAC-AF score could be used to identify a high-risk population that may be targeted for aggressive POAF preventive and screening measures, or to include in clinical trials of AF or stroke-preventive strategies following cardiac surgery.