Machine learning to risk stratify individuals for undiagnosed atrial fibrillation at scale using population-wide electronic health records.

AIMS: Electronic health records (EHR) can be used to target atrial fibrillation (AF) screening. We evaluated the performance of risk prediction models scalable across nationwide EHRs. METHODS: Retrospective cohort study individuals aged ≥30 years without diagnosed AF in the Clalit Health Services (Israel) EHR dataset between January 1 2019 and June 30, 2019. The primary outcome was a diagnosis of AF or atrial flutter (AFl) within 6 months. The FIND-AF, CHA2DS2-VASc and C2HEST scores were evaluated, with prediction performance assessed overall and by sex. The optimum threshold to apply in prospective screening was determined with a lift analysis. RESULTS: Of 2,166,795 individuals in the cohort 4,275 developed AF within 6 months. Prediction performance was strongest for FIND-AF (AUROC 0.871, 95% CI 0.864-0.877; calibration slope 0.73, 95% CI 0.67-0.79) compared with CHA2DS2-VASc (AUROC 0.838, 95% CI 0.831-0.845; calibration slope 0.63, 95% CI 0.60-0.67) and C2HEST scores (AUROC 0.834, 95% CI 0.823-0.844; calibration slope 0.62, 95% CI 0.58-0.65), including in women (FIND-AF AUROC 0.883, 95% CI 0.876-0.889; CHA2DS2-VASc AUROC 0.865, 95% CI 0.858-0.872; C2HEST AUROC 0.853, 95% CI 0.846-0.861) and men (FIND-AF AUROC 0.857, 95% CI 0.850-0.864; CHA2DS2-VASc AUROC 0.835, 95% CI 0.828-0.843; C2HEST AUROC 0.814, 95% CI, 0.806-0.822). Lift analysis suggested that screening the top 15% of FIND-AF risk compared to screening by age would identify 72% compared to 63% of AF diagnoses. CONCLUSION: The FIND-AF machine learning algorithm was scalable in routine EHR data with good discrimination for incident AF. Prospective evaluation is now required to evaluate risk-guided AF screening.