Diagnostic or prognostic study
Deep learning to detect left ventricular structural abnormalities in chest X-rays
- Diagnostic accuracy study
- external validation
- comparison with radiologists
Authors' conclusion
Deep learning analysis of CXRs can accurately detect the presence of certain structural abnormalities and may be useful in early identification of patients with LV hypertrophy and dilation.
Study context
- Population
- Patients with Chest X-rays (CXRs) completed within 1 year of echocardiograms
- Participants
- 24,689 patients · 71,589 CXRs
- Comparison
- Deep learning model on CXRs vs radiologists' visual assessment
- Radiologistsn = 15reference
- Outcome
- AUROC for SLVH, DLV and composite label (echocardiographic labels)
Performance
As reported in the abstract
AUROC for SLVH, DLV and composite label
Model AUROC, severe left ventricular hypertrophy (SLVH)
0.79 (0.76-0.81)
Model AUROC, dilated left ventricle (DLV)
0.80 (0.77-0.84)
Model AUROC, composite label
0.80 (0.78-0.83)
Sensitivity, model vs radiologists' consensus vote, at fixed specificity of 73%
71% vs. 66%
Further findings
| Outcome | Deep learning model | Radiologists |
|---|---|---|
| Composite label vs radiologists | ||
| Sensitivity at fixed specificity 73% | 71% | Not reported in the abstract |
| Sensitivity (consensus vote) at fixed specificity 73% | Not reported in the abstract | 66% |
External validation (8003 CXRs)
- Similar performance on an external data set
Composite label vs radiologists
- Model outperformed all 15 individual radiologists for the composite label
Original publication
Deep learning to detect left ventricular structural abnormalities in chest X-rays. Eur Heart J. Published 20 March 2024. doi:10.1093/eurheartj/ehad782
Summarised from the published abstract; every number is checked against it