Prediction of atrial fibrillation in patients without heart failure: validation study in 3.3 million individuals

The European Heart Rhythm Association Position Paper guidelines recommend searching for atrial fibrillation (AF) early before sequelae.1 Prediction models are widely used to identify individuals at high risk of incident AF;2 however, these models are often validated in cohorts including individuals with pre-existing heart failure HF,3 a strong risk factor for AF and a Sub-cohort of patients who may be under more intensive cardiac rhythm surveillance anyway.4

We aimed to evaluate the prediction performance of scalable prediction models for AF in patients without HF in a nationwide cohort.

We used UK primary care electronic health record (EHR) data from individuals aged ≥30 years without AF or HF the CPRD-AURUM dataset (Jan 1998 to Feb 2022). We assessed the performance of the FIND-AF, CHA₂DS₂-VASc, and C2HEST algorithms as these can be applied at scale without restriction by missing data in routine care. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) metrics.

Of 3 341 228 individuals included in the study (1 607 699 women (48.1%), mean age 62.46(SD 14.29), 2 571 344(76.95%) white ethnicity) (Table 1), 14 753 (0.44%) developed AF within 6 months. FIND-AF, CHA₂DS₂-VASc and C2HEST scores could be applied to all records. All 3 prediction models showed similar good prediction accuracy for incident AF (FIND-AF AUROC 0.7214(95% CI), CHADSVASC 0.7216(95% CI), C2HEST 0.719(95% CI) (Table 2). Low incidence of new-onset AF led to uniformly high NPV across algorithms. Sensitivity was highest for CHA₂DS₂-VASc (FIND-AF 0.619, CHADSVASC 0.7947) but specificity was highest for FIND-AF (FIND-AF 0.7072, CHADSVASC 0.5605).

FIND-AF, CHA₂DS₂-VASc and C2HEST showed good prediction performance for short-term incident AF in patients without HF in a nationwide dataset. Assessment in a prospectively screened cohort will provide more accurate information to the clinical utility of these models to target screening for AF in patients without known HF.Table 1  Table 2