Heart failure & cardiomyopathiesDiagnostic or prognostic study
Detection of Left Ventricular Systolic Dysfunction From Electrocardiographic Images
- Diagnostic accuracy study
- multicentre
- external validation
- prospective cohort validation
Authors' conclusion
We developed and externally validated a deep learning model that identifies LV systolic dysfunction from ECG images.
Study context
- Population
- Adults with 12-lead ECGs and echocardiograms within 15 days, Yale New Haven Hospital 2015–2021; external validation at US centres and the prospective ELSA-Brasil cohortECGs, model development
- Participants
- 385,601 adults
- Comparison
- Internal validation vs Cedars Sinai vs Outpatient Yale New Haven Hospital clinics vs Lake Regional Hospital vs Memorial Hermann Southeast Hospital vs Methodist Cardiology Clinic vs Brazilian Longitudinal Study of Adult Health
- Outcome
- Detection of LVEF <40% (AUROC, AUPRC)
Performance
As reported in the abstract
Detection of LVEF <40% from ECG images
| Test or model | AUROC | AUPRC |
|---|---|---|
| Internal validation | 0.91 | 0.55 |
| Cedars Sinai | 0.90 | 0.53 |
| Outpatient Yale New Haven Hospital clinics | 0.94 | 0.77 |
| Lake Regional Hospital | 0.90 | 0.88 |
| Memorial Hermann Southeast Hospital | 0.91 | 0.88 |
| Methodist Cardiology Clinic | 0.90 | 0.74 |
| Brazilian Longitudinal Study of Adult Health | 0.95 | 0.45 |
Further findings
Left ventricular (LV) systolic dysfunction on transthoracic echocardiogram
| Comparison | OR (95% CI) | |
|---|---|---|
| ECG suggestive of LV systolic dysfunction · held-out set | 27.5 (22.3–33.9) | |
| 0.050.21520Odds ratio (log scale) |
Incident LV systolic dysfunction (LVEF ≥40% at initial assessment) · median follow-up, 3.2 years
| Comparison | HR (95% CI) | |
|---|---|---|
| Positive ECG screen | 3.9 (3.3–4.7) | |
| 0.20.5125Hazard ratio (log scale) |
Class-discriminative patterns
- Localized to anterior and anteroseptal leads (V2 and V3), regardless of ECG layout
Original publication
Detection of Left Ventricular Systolic Dysfunction From Electrocardiographic Images. Circulation. Published 25 July 2023. doi:10.1161/circulationaha.122.062646
Summarised from the published abstract; every number is checked against it