Deciphering how macroscopic brain dynamics emerge from local neuronal interactions remains a fundamental challenge in systems neuroscience. While Functional Connectivity (FC) has successfully mapped statistical dependencies in the resting brain, it cannot fully explain the directed, causal mechanisms - Effective Connectivity (EC) - that generate these patterns. Current generative models capable of inferring EC are often limited by computational costs or a heavy reliance on anatomical priors, restricting their application to whole-brain Magnetoencephalography (MEG) analysis. This thesis introduces and validates the Recurrent Hopfield Mass Model (RHoMM), a novel, scalable, data-driven framework designed to estimate whole-brain EC from source-reconstructed MEG band-limited power signals. By modelling brain regions as binary McCulloch-Pitts neurons within an asymmetric recurrent network, RHoMM captures non-linear dynamics and causal influences. The model is trained via Backpropagation Through Time (BPTT), and extensive simulations demonstrate its ability to scale efficiently to networks of up to 200 nodes while accurately recovering ground-truth connectivity. When applied to resting-state MEG data, RHoMM faithfully reproduces empirical FC topologies. Crucially, the model uncovers a distinct architecture of communication: higher functional connectivity is driven by regions with large excitatory effective connections, particularly characterizing rich-club hubs within the Default Mode Network, while inhibitory connections support functional segregation in peripheral regions, such as the Visual Network. This dissociation is found to be most prominent in the alpha frequency band. Finally, a comparative analysis reveals that RHoMM could outperform the mechanistic Coupled Oscillator model in capturing the complex, non- linear dynamics of the resting brain without requiring structural connectivity constraints. These results establish RHoMM as a robust tool for investigating the causal wiring of the human brain.
A new generative model for MEG/EEG data: optimization, characterization and applications
FERRAZZA, MARTINA
2026-04-09
Abstract
Deciphering how macroscopic brain dynamics emerge from local neuronal interactions remains a fundamental challenge in systems neuroscience. While Functional Connectivity (FC) has successfully mapped statistical dependencies in the resting brain, it cannot fully explain the directed, causal mechanisms - Effective Connectivity (EC) - that generate these patterns. Current generative models capable of inferring EC are often limited by computational costs or a heavy reliance on anatomical priors, restricting their application to whole-brain Magnetoencephalography (MEG) analysis. This thesis introduces and validates the Recurrent Hopfield Mass Model (RHoMM), a novel, scalable, data-driven framework designed to estimate whole-brain EC from source-reconstructed MEG band-limited power signals. By modelling brain regions as binary McCulloch-Pitts neurons within an asymmetric recurrent network, RHoMM captures non-linear dynamics and causal influences. The model is trained via Backpropagation Through Time (BPTT), and extensive simulations demonstrate its ability to scale efficiently to networks of up to 200 nodes while accurately recovering ground-truth connectivity. When applied to resting-state MEG data, RHoMM faithfully reproduces empirical FC topologies. Crucially, the model uncovers a distinct architecture of communication: higher functional connectivity is driven by regions with large excitatory effective connections, particularly characterizing rich-club hubs within the Default Mode Network, while inhibitory connections support functional segregation in peripheral regions, such as the Visual Network. This dissociation is found to be most prominent in the alpha frequency band. Finally, a comparative analysis reveals that RHoMM could outperform the mechanistic Coupled Oscillator model in capturing the complex, non- linear dynamics of the resting brain without requiring structural connectivity constraints. These results establish RHoMM as a robust tool for investigating the causal wiring of the human brain.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
