Concept Bottleneck Models Tested on Echocardiography Volume Grounding
A new arXiv preprint (2607.25748) investigates concept bottleneck models in echocardiography. Researchers trained a video transformer encoder on a public dataset to predict left ventricular volumes as intermediate concepts for ejection fraction estimation. They compared training with ejection fraction objective alone versus additional supervision of volumes in milliliters. The concept bottleneck did not increase error (6.89 vs 7.13 MAE). The study questions whether accuracy of concept predictions alone validates model interpretability.
Key facts
- arXiv preprint 2607.25748
- Concept bottleneck models route prediction through interpretable intermediate variables
- Left ventricular volumes used as concepts for ejection fraction estimation
- Video transformer encoder trained on public echocardiography dataset
- End-systolic and end-diastolic volumes formed concept layer
- Ejection fraction computed analytically with no residual path
- Training compared: ejection fraction objective alone vs additional volume supervision
- Evaluated on 1276 held-out studies
- Concept bottleneck: 6.89 MAE, direct regression: 7.13 MAE
Entities
Institutions
- arXiv