Heart failure & cardiomyopathies
Deep Learning to Estimate Left Ventricular Ejection Fraction From Routine Coronary Angiographic Images
Authors' conclusion
A novel DL algorithm demonstrated rapid and accurate assessment of LVEF from routine coronary angiography.
Study context
- Population
- Patients undergoing diagnostic coronary angiography at Mayo Clinic, with echocardiogram within 30 days as gold standard for left ventricular ejection fraction (LVEF)18,809 angiograms
- Participants
- 17,346 patients
- Comparison
- Deep learning (DL) model vs Human expert
- Outcome
- Classification of LVEF as significantly reduced (≤40%) vs normal or mildly reduced (>40%)
Methodological detailsdeep learning (combined 3D-CNN and transformer), training/validation/testing split (70%/10%/20%)
- Deep learning (combined 3D-CNN and transformer)
- Training/validation/testing split (70%/10%/20%)
Performance
As reported in the abstract
Detection of LVEF ≤40%
| Test or model | AUC | Sensitivity | Specificity |
|---|---|---|---|
| DL model | 0.86 | 0.76 | 0.77 |
| Human expert | 0.76-0.77 | 0.50-0.44 | 0.90-0.93 |
AUC
DL model0.86
Human expert0.76-0.77
Sensitivity
DL model0.76
Human expert0.50-0.44
Specificity
DL model0.77
Human expert0.90-0.93
Further findings
Model performance in training set
- AUC
- 0.87
- Sensitivity
- 0.77
- Specificity
- 0.80
Combined left anterior oblique (LAO) and right anterior oblique (RAO) views vs either view alone
- Combined LAO and RAO views yielded higher AUC, sensitivity and specificity than LAO or RAO individually
- Sensitivity analyses
Combined CNN and transformer vs 3D-CNN or transformer alone
- Combined CNN and transformer model superior to DL models using either 3D-CNN or transformers
Original publication
Deep Learning to Estimate Left Ventricular Ejection Fraction From Routine Coronary Angiographic Images. JACC Adv. Published 11 October 2023. doi:10.
Summarised from the published abstract; every number is checked against it