Multi-channel causal variational autoencoder for multimodal biomedical causal disentanglement

The multimodal nature of clinical assessment and decision-making, and the high rate of healthcare data generation, motivate the need to develop approaches specifically tailored to the analysis of these complex and potentially high-dimensional multimodal datasets. This poses both technical and conceptual challenges: how can such heterogeneous data be analyzed jointly? How can modality-specific information be identified from shared information? Variational autoencoders (VAEs) offer a robust framework for learning latent representations of complex data distributions, while being flexible enough to adapt to different data types and structures, and having already been successfully applied for latent disentanglement of multimodal (multi-channel) data. We aim at tackling multi-channel disentanglement from a causal perspective, and seek at identifying causal relationships between channels, beyond simple statistical associations. To do that, we propose Multi-Channel Causal VAE (MC²VAE), a novel causal disentanglement approach for multi-channel data, whose objective is to jointly learn modality-specific latent representations from a multi-channel dataset, and identify a causal structure between the latent channels. Each channel is projected into its own latent space, where a causal discovery step is integrated to learn the hidden causal graph. Finally, the decoder takes into account the discovered graph to predict the data. Covariate of interest can be integrated as well when available, and accounted in the causal graph structure. Extensive experiments on synthetically generated multi-channel datasets demonstrate the ability of MC²VAE in effectively uncovering the underlying latent causal structures across multiple channels, hence making it a strong candidate for real-world multi-channel causal disentanglement. Application to multi-channel data on neurodegeneration extracted from the Alzheimer's Disease Neuroimaging Initiative highlights the existence of a biologically meaningful latent causal structure, whose pertinence is supported by multiple previous experimental and modeling work, and provides actionable insight for disease progression.