Multimodal Representation for Better Genetic Discovery from Physiological Waveforms

Yuchen Zhou will discuss a novel method that applies a multimodal AI model to physiological waveforms, like ECGs, to improve the genetic prediction of cardiovascular traits. The presentation will cover how this novel approach leveraged the shared and orthogonal signals in multimodal data and led to better genetic discovery.

Overview of Presentation

• This study investigated if deep learning representations of multimodal physiological waveforms (e.g., ECG, PPG) could enhance genomic discovery for cardiovascular traits.
• Variational Autoencoder (VAE) models trained on multimodal medical waveform data jointly and independently were developed to validate the advantage of multimodal learning.
• Genome-Wide Association Studies (GWAS) were then performed on these AI-generated latent embeddings from multimodal VAE and Unimodal VAE to search for genetic associations.
• The approach successfully identified novel genetic loci associated with heart function, which were not discoverable using more traditional GWAS methods.
• These findings demonstrate that integrating multimodal AI with large-scale genetic data can significantly improve the power of genomic discovery and enhance the prediction of complex traits.