Incidence and severity of aortic stenosis according to machine learning predicted risk of atrial fibrillation

Abstract

Abstract Atrial fibrillation (AF) and aortic stenosis (AS) are two common progressive conditions affecting older persons that share pathobiological pathways. Early detection of AS is critical for improving outcomes, but no prediction tool exists to inform decision making. In this study we evaluated the association between machine learning predicted risk of incident AF from clinical health records (using the FIND-AF algorithm) and severity and incidence of AS. In a disease registry we found that higher FIND-AF risk was correlated with parameters of increasing AS severity including smaller aortic valve area, and higher maximum velocity and peak pressure gradient but ability to differentiate severe from non-severe AS was moderate (sensitivity 0.545, specificity 0.770). In over 400,000 primary care clinical health records, FIND-AF showed good prediction performance for incident AS (AUC 0.782, 95% CI 07.69–0.795), and the cumulative incidence increased with higher FIND-AF risk strata. The hazard of AS was over 40-fold higher in patients with FIND-AF risk scores of more than 0.05 compared to patients with FIND-AF risk scores of less than 0.005. Predicted risk of AF is associated with severity and incidence of AS, but predictive ability for AS may be improved by developing a machine learning model specifically for this outcome.