Estimating a brain network predictive of stress and genotype with supervised autoencoders

Abstract

Targeted brain stimulation has the potential to treat mental illnesses. We develop an approach to help design protocols by identifying relevant multi-region electrical dynamics. Our approach models these dynamics as a superposition of latent networks, where the latent variables predict a relevant outcome. We use supervised autoencoders (SAEs) to improve predictive performance in this context, describe the conditions where SAEs improve predictions, and provide modelling constraints to ensure biological relevance. We experimentally validate our approach by finding a network associated with stress that aligns with a previous stimulation protocol and characterizing a genotype associated with bipolar disorder.

Electome network factors: Capturing emotional brain networks related to health and disease

“…13 More recently, we developed a supervised autoencoder approach to learn the electome networks, a deep learning technique common in modern machine learning that uses information from the associated behavior to force the supervised electome to capture relevant brain dynamics, even when this electome network explains a relatively small percentage of the total electrophysiological signal. 47 We demonstrated that this technique is a promising approach to address multiple sources of misspecification. 47 While unsupervised approaches require less effort to train models, they often find the strongest neural signal or the neural features that explain the most variance, not those that may be most relevant to the topic of interest.…”

“… 47 We demonstrated that this technique is a promising approach to address multiple sources of misspecification. 47 While unsupervised approaches require less effort to train models, they often find the strongest neural signal or the neural features that explain the most variance, not those that may be most relevant to the topic of interest.…”