Machine Learning to Infer and Control Brain State
It is increasingly possible to develop treatments for psychiatric disorders by making targeted interventions on the brain. However, designing an appropriate protocol requires many choices. We propose a method that identifies electrical dynamics across brain regions related to illness states or behaviors and employs these patterns to design intervention protocols. Specifically, the observed electrical activity of the brain is statistically modeled as a superposition of activity from latent electrical functional connectome (electome) networks. The activity of these latent networks defines a brain state that predicts disease state, behavior, or outcomes. These electome networks are explainable in their spectral power and directional relationships between brain regions, facilitating the design of testable protocols on key relationships. We present a case study on social aggression, where we identify an electome network associated with aggressive behavior and develop a machine-learning controlled protocol that selectively reduces aggression without affecting pro-social behavior. We conclude with ongoing efforts in causal discovery and mediation analysis to further understand and improve this system.