Modelling discrete states and long-term dynamics in functional brain networks.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Functional brain network dynamics are crucial for cognition, behavior, memory, aging, and clinical disorders.
• Unsupervised machine learning can infer brain network dynamics from electrophysiological data at sub-second timescales.
• Existing methods like Hidden Markov Models (HMMs) offer interpretability but lack long-range temporal modeling, while deep learning models capture temporal structure at the cost of interpretability.

Purpose of the Study:

• To introduce Dynamic Network States (DyNeStE), a novel computational framework.
• To address the trade-off between interpretability and temporal modeling in brain network dynamics analysis.
• To develop a model that captures both categorical brain states and long-range temporal dependencies.

Main Methods:

• DyNeStE utilizes amortised Bayesian inference with recurrent neural networks.
• A Gumbel-Softmax distribution is employed to enforce categorical states for enhanced interpretability.
• The model was evaluated using both simulated data and real resting-state magnetoencephalography (MEG) data.

Main Results:

• DyNeStE successfully recovered plausible dynamic brain network states in simulations and real MEG data.
• The model demonstrated superior performance over HMM in capturing long-range temporal dependencies.
• Recovered dynamic networks were reproducible across independent data splits and aligned with existing HMM findings.

Conclusions:

• DyNeStE provides an interpretable and temporally informative framework for analyzing large-scale neural activity.
• The model effectively represents neural activity as discrete state transitions.
• DyNeStE captures transient and long-range brain network dynamics, offering advancements over existing methods.