Arrhythmias & electrophysiologyDiagnostic or prognostic study
Prediction of incident atrial fibrillation using deep learning, clinical models, and polygenic scores
- Prognostic model development and validation study
- single-centre development (tertiary cardiac centre)
- 70/10/20 training/validation/test split
- external validation (Medical Information Mart for Intensive Care-IV (MIMIC-IV))
Authors' conclusion
ECG-AI provides an accurate tool to predict new-onset AF in a tertiary cardiac centre, surpassing clinical and PGS.
Study context
- Population
- Patients with sinus-rhythm ECGs at the Montreal Heart Institute without pre-existing AF
- Participants
- 145,323 patients · 669,782 ECGs
- Comparison
- ECG-AI vs clinical models and AF polygenic score
- Outcome
- Incident AF
- Follow-up
- 5 years
Performance
As reported in the abstract
Incident AF at 5 years (observed in 15%)
| Test or model | Value | AUC |
|---|---|---|
| ECG-AI | .78 | .78 |
- Primary outcome observed in
- 15%
Further findings
| Outcome | ECG-AI | CHARGE-AF | PGS | fold increased risk (95% CI) |
|---|---|---|---|---|
| Incident AF with ECG-AI high-risk inference (time-to-event, first ECG) · high risk = 26% of population | 4.3 (4.02–4.57) | |||
| Incident AF, subgroup of 2301 patients | ||||
| AUC | Not reported in the abstract | .62 | .59 | |
| External validation (MIMIC-IV) · median follow-up 1.1 year | ||||
| AUC | .77 | Not reported in the abstract | Not reported in the abstract |
- Incident AF, subgroup of 2301 patients — Adding polygenic score (PGS) and CHARGE-AF to ECG-AI improved goodness of fit (likelihood ratio test)
- P < .001
- Incident AF, subgroup of 2301 patients — AUC with added PGS and CHARGE-AF (minimal change)
- .76-.77
- External validation (MIMIC-IV): mean age 59 ± 18 years, 47% male
Incident AF, subgroup of 2301 patients
- ECG-AI outperformed CHARGE-AF and PGS
Original publication
Prediction of incident atrial fibrillation using deep learning, clinical models, and polygenic scores. Eur Heart J. Published 1 September 2024. doi:10.1093/eurheartj/ehae595
Summarised from the published abstract; every number is checked against it