Modelling variability in dynamic functional brain networks using embeddings

Abstract

AbstractNeuroimaging techniques offer unprecedented insights into the dynamic neural processes underlying cognitive functions. With recent studies, data driven models like the Hidden Markov Model (HMM) are getting more attention due to their ability to infer fast temporal dynamics in functional networks in an unsupervised manner. However, these dynamic network models are typically trained at the group level. Whilst it is possible to post-hoc estimate the session-specific networks with the so-called dual estimation, this does not allow the model to discover and benefit from subpopulation structure in the group. We propose an extension to the HMM model that incorporates embedding vectors (c.f. word embedding in Natural Language Processing) to explicitly model individual sessions while training on the entire group. This effectively infers a “fingerprint” for each individual session, which can group together those with similar spatio-temporal patterns. With simulated data, we show that the model can recover the underlying subpopulation structure, achieve higher accuracy than dual estimation on session-specific quantities and can make use of increasing number of sessions to benefit the inference of individual sessions. Applying this model to resting-state and task MEG data, we show the learnt embedding vectors capture meaningful sources of variation across a population. This includes subpopulations related to demographics and systematic differences, such as scanner types or measurement sites. The proposed model offers a powerful new technique for modelling individual sessions while leveraging information from an entire group.