Characterizing Continuous and Discrete Hybrid Latent Spaces for Structural Connectomes

Area of Science:

• Neuroscience
• Machine Learning
• Biostatistics

Background:

• Structural connectomes map brain connectivity but are high-dimensional and difficult to analyze.
• Existing low-dimensional methods like PCA and standard autoencoders struggle to model mixed continuous and discrete variability in connectomes.
• Understanding variability sources is crucial for aging, cognition, and neurodegenerative disease research.

Purpose of the Study:

• To develop and evaluate a variational autoencoder (VAE) with a hybrid latent space capable of jointly modeling discrete and continuous variability in structural connectomes.
• To assess the VAE's ability to disentangle different sources of variation, such as imaging site effects, in large-scale connectome datasets.
• To demonstrate the potential of this hybrid approach for improved interpretability and analysis of high-dimensional brain network data.

Main Methods:

• Proposed a variational autoencoder (VAE) with a hybrid latent space designed to simultaneously model discrete and continuous factors within connectome data.
• Analyzed a large dataset comprising 5,761 structural connectomes from 6 Alzheimer's disease studies, encompassing diverse demographics and cognitive statuses.
• Trained the hybrid VAE in an unsupervised manner and evaluated its performance in capturing and separating sources of variability, particularly site-specific differences.

Main Results:

• The hybrid VAE successfully modeled both discrete and continuous components of variability in the connectome data.
• The discrete component of the latent space effectively captured site-related differences, achieving an Adjusted Rand Index (ARI) of 0.65.
• This performance significantly surpassed traditional methods like PCA and standard VAE followed by clustering (p << 0.05), highlighting the hybrid model's superiority in disentangling variability.

Conclusions:

• The proposed hybrid latent space VAE offers a powerful unsupervised method for disentangling distinct sources of variability in structural connectomes.
• This approach demonstrates significant advantages over conventional techniques in identifying subtle factors like imaging site effects.
• The findings suggest promising potential for large-scale connectome analysis, aiding research in aging, cognition, and neurodegenerative diseases.