Development and validation of a novel machine learning-based algorithm to predict incident atrial fibrillation: A...

Existing atrial fibrillation (AF) risk prediction models incorporate race as a covariate, systematically underestimating AF risk in black individuals and potentially perpetuating health care disparities.

This study aimed to develop and validate machine learning (ML)-based race-agnostic risk scores to predict AF risk and assess differences in risk stratification and bias compared with the CHARGE-AF score.

The derivation cohort included 16,719 participants free of AF at baseline (Atherosclerosis Risk in Communities visit 5, 2011-2013; Cardiovascular Health Study baseline, 1989-1990), and the validation cohort included 13,928 (Multi-Ethnic Study of Atherosclerosis and Framingham Offspring and Generation 3 studies). The primary outcome was the incidence of AF within 5 years. Model performance was assessed using concordance index, Brier score, and index of prediction accuracy. Bias was evaluated using disparate impact, equal opportunity difference, and Theil index. Population-attributable risk percentage was calculated across racial groups.

During the 5-year follow-up, incident AF occurred in 507 participants (3.0%) in the derivation cohort and 262 (1.9%) in the validation cohort. The ML model demonstrated superior performance compared with CHARGE-AF, with better discrimination (concordance index 0.83 [95% confidence interval 0.80-0.85] vs 0.77 [95% confidence interval 0.74-0.79]; P < .001) and improved calibration (Brier score 1.82 vs 1.92; P < .001). Key predictors included age, clinical factors (electrocardiographic parameters, cardiac biomarkers, and blood pressure), and education level. Population-attributable risk analysis demonstrated marked racial differences in AF risk contribution from age (non-Hispanic black 14.3% vs white participants 34.6%). The ML model reduced algorithmic bias vs CHARGE-AF across all metrics.

Race-agnostic ML models demonstrated superior predictive performance and reduced bias compared with CHARGE-AF, potentially improving clinical risk stratification while promoting health equity.