Multi-channel causal variational autoencoder for multimodal biomedical causal disentanglement.

Area of Science:

• Computational biology
• Machine learning
• Causal inference

Background:

• Healthcare data is increasingly multimodal and high-dimensional, posing challenges for analysis.
• Variational autoencoders (VAEs) are effective for learning latent representations and disentangling multimodal data.
• Existing methods often focus on statistical associations rather than causal relationships.

Purpose of the Study:

• To develop a novel approach for causal disentanglement of multi-channel healthcare data.
• To identify modality-specific latent representations and causal structures within heterogeneous datasets.
• To apply the method to neurodegeneration data for insights into disease progression.

Main Methods:

• Proposed Multi-Channel Causal Variational Autoencoder (MC2VAE) for joint learning of latent representations and causal structure.
• Integrated causal discovery within each channel's latent space to learn a hidden causal graph.
• Decoder incorporates the discovered causal graph for data prediction.
• Option to integrate covariates of interest into the causal graph.

Main Results:

• MC2VAE effectively uncovers underlying latent causal structures in synthetic multi-channel datasets.
• Demonstrated ability to identify biologically meaningful causal structures in Alzheimer's Disease Neuroimaging Initiative data.
• Provided actionable insights into neurodegeneration disease progression.

Conclusions:

• MC2VAE is a robust framework for causal disentanglement of multi-channel data.
• The approach offers a powerful tool for analyzing complex biomedical datasets.
• Identified causal structures can inform understanding of disease mechanisms and progression.