A VQ-VAE framework for modeling physiological information in fMRI

The coupling between functional magnetic resonance imaging (fMRI) and autonomic physiological processes is not merely “noise”, but captures important brain-body interactions. This relationship is likely bidirectional: autonomic fluctuations can influence brain activity and hemodynamics, and neuronal activity can modulate systemic physiological processes. This complexity makes it difficult for traditional linear approaches to fully characterize their relationship. Here, we introduce a novel application of vector quantized variational autoencoder (VQ-VAE) to characterize whole-brain patterns associated with respiration using the discretized fMRI latent space. Further, we demonstrate that this framework can be leveraged to assess the quality of fMRI-based reconstructions of low-frequency respiratory fluctuations when physiological recordings are missing or corrupted. The success of our model indicates that we can extract the non-linear relationship between the fMRI brain signals and bodily autonomic processes, revealing their intrinsic connections in the cross-modal latent space.