Multi-trait polygenic risk scores improve genomic prediction of atrial fibrillation across diverse ancestries

Abstract Polygenic scores can improve atrial fibrillation risk prediction. However, limited accuracy and cross-ancestry transferability hinder clinical translation. Here, we explore several ensemble approaches to generate ancestry-optimized polygenic scores, with development in diverse participants from the All of Us Research Program, BioBank Japan, and three additional cohorts. Our ancestry-specific multi-trait approach particularly improves prediction in South-Asian (odds-ratio/standard deviation 1.5–1.8; area under curve 0.60-0.64; relative R² +71%), Admixed-American (1.5; 0.60; +34%) and African ancestry groups (1.4; 0.57; +56%). Nevertheless, performance remains highest in European and East-Asian ancestries (1.8–2.2; 0.65–0.68), where >50% of SNP-heritability is explained. Improved risk stratification is also observed at the extremes, identifying European and East-Asian ancestry individuals with risk comparable to rare TTN variants (e.g., 6–11% with >4-fold odds). Finally, our scores improve incident risk prediction alongside clinical models. Together, we show that our ancestry-tailored multi-trait polygenic scores advance atrial fibrillation risk prediction and stratification, providing an equitable foundation for implementation.