Mixtures of large-scale dynamic functional brain network modes.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Accurate temporal modeling of functional brain networks is crucial for understanding cognition.
• Current time-varying analyses for electrophysiological data often assume network mutual exclusivity, potentially limiting their effectiveness.
• Dynamic processes in the brain occur on millisecond timescales, requiring sophisticated modeling approaches.

Purpose of the Study:

• To propose a novel generative model, Dynamic Network Modes (DyNeMo), for functional brain connectivity.
• To enable the modeling of time-varying functional connectivity as a mixture of dynamic network modes.
• To utilize a recurrent neural network and amortised variational inference for parameter learning.

Main Methods:

• Developed DyNeMo, a generative model representing functional connectivity as a time-varying linear mixture of statistical modes.
• Governed temporal evolution of the mixture using a recurrent neural network.
• Employed amortised variational inference within a Bayesian framework to learn model parameters.

Main Results:

• DyNeMo outperforms Hidden Markov Modelling (HMM) when the assumption of mutual exclusivity is violated, as shown in simulations.
• In resting-state MEG data, DyNeMo identified modes activating on fast timescales (100-150 ms).
• DyNeMo detected plausible, task-dependent evoked responses in task MEG data without prior knowledge of task timings.

Conclusions:

• DyNeMo offers a flexible framework for modeling dynamic brain networks and assessing future developments.
• The model provides decompositions that approximate HMM remappings but with improved explanatory power.
• While HMM's mutual exclusivity assumption can be reasonable, DyNeMo offers a more nuanced approach to capturing brain dynamics.