Valvular & structural heart diseaseDiagnostic or prognostic study
Severe aortic stenosis detection by deep learning applied to echocardiography
- Diagnostic accuracy study
- deep learning model development and validation
- temporal and geographic external validation
Authors' conclusion
This study developed and externally validated an automated approach for severe AS detection using single-view 2D echocardiography, with potential utility for point-of-care screening.
Study context
- Population
- Transthoracic echocardiography studies (2D parasternal long axis videos without Doppler) for detection of severe aortic stenosis (AS)training set 5257 studies
- Participants
- 2,040 / 4,226 / 3,072
- Comparison
- Temporally distinct test set vs California vs New England
- Temporally distinct test setn = 2,040
- Californian = 4,226
- New Englandn = 3,072
- Outcome
- AUROC for detecting severe AS
Performance
As reported in the abstract
Severe AS detection
| Test or model | AUROC |
|---|---|
| Yale-New Haven Hospital (YNHH) temporal test set | 0.978 (95% CI: 0.966, 0.988) |
| California | 0.952 (95% CI: 0.941, 0.963) |
| New England | 0.942 (95% CI: 0.909, 0.966) |
Further findings
Interpretability (saliency maps)
- Saliency maps identified the aortic valve, mitral annulus and left atrium as predictive regions
Non-severe AS cases
- Predicted probabilities associated with worse quantitative metrics of AS
Original publication
Severe aortic stenosis detection by deep learning applied to echocardiography. Eur Heart J. Published 23 August 2023. doi:10.1093/eurheartj/ehad456
Summarised from the published abstract; every number is checked against it