Heart failure & cardiomyopathies
Algorithmic Guardrails for Echocardiography Artificial Intelligence (AI)
- retrospective validation
Authors' conclusion
Deep feature representations can be leveraged to identify and automatically flag misinterpretation-prone echocardiograms, enabling promising guardrail mechanisms for the reliable deployment of echocardiography AI.
Study context
- Population
- Apical-4-chamber echocardiography videos with biplane left ventricular ejection fraction (LVEF) measurements in the MIMIC-IV-ECHO database (2017–2019)4,010 apical-4-chamber videos from 1,353 studies
- Participants
- 1,208 participants
- Comparison
- EchoNet minus PanEcho
Performance
As reported in the abstract
LVEF estimation performance (mean absolute error, R2)
EchoNet vs PanEchoΔmean absolute error (CI)
EchoNet minus PanEcho; PanEcho performed better
0.31(0.07–0.55)
- EchoNet minus PanEcho
- ΔR2 −0.07 (CI −0.11 to −0.03)
Further findings
| Outcome | EchoNet | PanEcho |
|---|---|---|
| Proportional bias slope (systematic overestimation of LVEF at lower values) | proportional bias slope −0.33 (interval −0.39–-0.28) | proportional bias slope −0.31 (interval −0.36–-0.26) |
| P | P < 0.001 | P < 0.001 |
| Selective estimation using intermediate features from EchoNet (area under risk-coverage curve) | area under risk-coverage curve 5.04 (interval 4.45–5.66) | area under risk-coverage curve 5.21 (interval 4.54–5.92) |
Error slicing by deep feature representations
- All investigated deep feature representations identified error slices for both models
- Shared error slices revealed 2 common failure modes: contrast-enhanced echocardiograms and poor-quality apical-4-chamber views
Selective estimation using intermediate features from EchoNet (area under risk-coverage curve)
- Best performance across both estimation models
Original publication
Algorithmic Guardrails for Echocardiography AI. JACC Adv. Published 13 October 2026. doi:10.
Summarised from the published abstract; every number is checked against it