Transferability of Structural ECG Representations Under Domain Shift

Determine whether structural representations learned by aligning electrocardiograms with cardiac magnetic resonance imaging transfer under domain shift and retain diagnostic value for diseases whose patients and imaging are entirely absent from the pre-training data.

Background

Multimodal ECG–CMR representation-learning methods aim to encode cardiac structural and functional information in ECG representations by training on paired electrocardiograms and cardiac magnetic resonance examinations. Prior evaluations have primarily used populations and disease distributions resembling those in the pre-training cohorts, leaving generalization to substantially different settings insufficiently established.

The unresolved issue is whether such structural representations remain diagnostically useful when applied to diseases and patient populations that were entirely absent from pre-training. The paper evaluates this question for Chagas disease, but the broader transferability of the learned representations under domain shift remains unresolved.

References

It therefore remains unknown whether the structural representations transfer under domain shift, and in particular, whether they retain diagnostic value for diseases whose patients and imaging are entirely absent from the pre-training data.

— Leveraging Cardiac Imaging to Improve ECG-Based Detection of Chagas Disease in Resource-Constrained Settings  (2609.08582 - Alvarez-Florez et al., 8 Sep 2026) in Section Introduction