External validation of atrial fibrillation risk scores in heart failure under continuous device surveillance

Population-derived atrial fibrillation (AF) risk scores are widely used, but their transportability to populations with heart failure (HF) under continuous rhythm surveillance is uncertain.

The goals of this study were to externally validate established AF risk scores in patients with HF monitored continuously using implantable devices and to assess discrimination and clinical utility for guiding surveillance strategies.

This study included a prospective multicenter cohort of outpatients with HF without prior AF. All patients carried implantable loop recorders or cardiac implantable electronic devices enabling continuous monitoring. Patients with pre-baseline atrial high-rate episodes lasting ≥6 minutes were excluded. Incident AF (≥6 minutes by device diagnostics or 12-lead electrocardiography) was the primary end point. We calculated 13 published AF risk scores as originally specified and evaluated discrimination (area under the receiver operating characteristic curve [AUC]), overall accuracy (Brier score), competing risk cumulative incidence, and 5-year decision curve net benefit.

Of the 396 patients, 98 (24.8%) developed AF during follow-up. FIND-AF (Future Innovations in Novel Detection of Atrial Fibrillation) score showed the highest discrimination (AUC 0.72 95% confidence interval 0.66-0.78), whereas most other scores performed modestly (typical AUC ≈0.50-0.61). Calibration was generally suboptimal with risk overestimation and slopes <1; FIND-AF required recalibration (intercept and slope adjustments) to align predicted and observed risks. At 5 years, decision curve analysis favored FIND-AF and MVP-ECG (Morphology-Voltage-P-wave Duration Electrocardiographic score), which provided the most consistent net benefit vs "treat all/none" across plausible threshold probabilities, although absolute gains were moderate.

In continuously monitored populations with AF, most population AF scores underperformed; FIND-AF performs best but remains below thresholds for stand-alone decisions. Routine reliance on population scores to guide surveillance in HF is not supported.