Predicting atrial fibrillation in heart failure: added value of a plasma protein risk score over clinical and genomic models

Abstract Introduction Atrial fibrillation (AF) commonly coexists with heart failure (HF), aggravating disease progression and prognosis. Early diagnosis and treatment of AF in heart failure (HF) is crucial as the two conditions mutually reinforce each other resulting in adverse outcome and prognosis. We aimed to develop and externally validate a plasma protein-based risk score for incident AF in HF, benchmarking its predictive performance against established clinical and genomic models. Methods Data from the MyoVasc HF cohort (n=3,289) were analyzed. Incident AF over 4 years was identified, via examination in the study center at follow-up visits. 536 proteins were analyzed using proximity extension assay technology (Olink®). After exclusion of BNP and NT-proBNP, proteins associated with incident AF were selected via elastic net-regularized logistic regression to derive the AF-protein-signature and weighted AF-protein-score. Predictive performance was benchmarked against CHARGE-AF and AF-PRS using multivariable robust Poisson regression in the deviation cohort (sex, age, cardiovascular risk factors (CVRFs), comorbidities, NT-proBNP, left ventricular ejection fraction). External validation was performed using UK Biobank Resource under application number 315074. Results Incidence of AF in the study sample was 6 % (n=126; 69 % men; mean age 68.6 ± 9.5 years). 18 AF-associated proteins were selected. The AF protein score predicted incident AF independent of the clinical profile (fully adjusted model, prevalence ratio Standard Deviation (PRSD) 1.69, 95% confidence interval (CI): 1.40; 2.1, p0.0001, AUC 0.80). Using single models adjusted for age and sex, discrimination of 4-year incident AF was highest for the AF protein score (PRSD = 1.99, 95% CI: 1.69; 2.33, p0.0001, AUC = 0.74), followed by CHARGE-AF (PRSD = 1.37, 95% CI: 1.18; 1.59, p0.0001, AUC = 0.69). The AF polygenic risk score (AF-PRS) showed no significant association with incident AF (PRSD = 1.04, 95% CI: 0.89; 1.22, p=0.6) and demonstrated lower discriminative ability (AUC = 0.66). In the validation cohort (n=41,049), after adjustment for age, sex, and CVRFs, the incident AF protein score showed the strongest association with AF risk (PRSD = 1.42, 95% CI 1.33 - 1.51) versus CHARGE-AF (PRSD = 1.33, 95% CI 1.27 - 1.39), with comparable discrimination (AUC 0.76 vs. 0.78). Benchmarking against the AF-PRS will be available at the congress. Conclusion A novel machine learning-derived score incorporating 18 circulation proteins demonstrates superior predictive performance for 4-year incident AF risk compared with established clinical and genetic models in patients with HF. Validation in an independent population-based cohort confirmed the robustness of these findings. Protein-based AF risk estimation may enhance the efficiency of AF screening and guide preventive interventions in populations at risk.