Keywords: multimodal learning, cardiology, structural heart disease, deep learning
Abstract: Structural heart disease (SHD) is typically diagnosed using transthoracic echocardiograms (TTEs), a modality underutilized in the United States. We investigate what combination of common clinical modalities in electrocardiograms (ECGs), posteroanterior view chest X-rays, and structured electronic health record (EHR) data can detect SHD labels generated with an TTE unseen by the model. Our experiments show that ECG-based models in both unimodal and multimodal settings performed best and that the inclusion of additional modalities with a late-fusion approach can give a marginal performance improvement.
Submission Number: 155
Loading