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

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence in Healthcare
• Health Equity Research

Background:

• Existing atrial fibrillation (AF) risk prediction models often include race, leading to systematic underestimation of risk in Black individuals and exacerbating healthcare disparities.
• This bias highlights the need for more equitable risk assessment tools in cardiovascular medicine.

Purpose of the Study:

• To develop and validate novel machine learning (ML)-based, race-agnostic risk scores for predicting AF incidence.
• To compare the performance and bias of these ML models against the established CHARGE-AF score.

Main Methods:

• Utilized large derivation and validation cohorts (ATHEROSCLEROSIS RISK IN COMMUNITIES, CARDIOVASCULAR HEALTH STUDY, MULTI-ETHNIC STUDY OF ATHEROSCLEROSIS, FRAMINGHAM OFFSPRING AND GENERATION 3) comprising over 30,000 participants.
• Assessed model performance using concordance index and Brier score; evaluated bias using disparate impact, equal opportunity difference, and Theil index.
• Identified key predictors including age, clinical factors (ECG, biomarkers, blood pressure), and education level.

Main Results:

• The ML models demonstrated superior predictive performance compared to CHARGE-AF, evidenced by higher concordance index (0.83 vs 0.77) and improved calibration (lower Brier score).
• Population-attributable risk analysis revealed significant racial differences in AF risk contribution from age.
• The developed ML models significantly reduced algorithmic bias compared to CHARGE-AF across all evaluated metrics.

Conclusions:

• Race-agnostic ML models offer enhanced predictive accuracy and calibration for AF risk assessment.
• These models effectively reduce bias, presenting a promising approach to improve clinical risk stratification and promote health equity in cardiovascular care.