A graph neural network framework for causal inference in brain networks

Abstract

A central question in neuroscience is how self-organizing dynamic interactions in the brain emerge on their relatively static structural backbone. Due to the complexity of spatial and temporal dependencies between different brain areas, fully comprehending the interplay between structure and function is still challenging and an area of intense research. In this paper we present a graph neural network (GNN) framework, to describe functional interactions based on the structural anatomical layout. A GNN allows us to process graph-structured spatio-temporal signals, providing a possibility to combine structural information derived from diffusion tensor imaging (DTI) with temporal neural activity profiles, like that observed in functional magnetic resonance imaging (fMRI). Moreover, dynamic interactions between different brain regions discovered by this data-driven approach can provide a multi-modal measure of causal connectivity strength. We assess the proposed model’s accuracy by evaluating its capabilities to replicate empirically observed neural activation profiles, and compare the performance to those of a vector auto regression (VAR), like that typically used in Granger causality. We show that GNNs are able to capture long-term dependencies in data and also computationally scale up to the analysis of large-scale networks. Finally we confirm that features learned by a GNN can generalize across MRI scanner types and acquisition protocols, by demonstrating that the performance on small datasets can be improved by pre-training the GNN on data from an earlier study. We conclude that the proposed multi-modal GNN framework can provide a novel perspective on the structure-function relationship in the brain. Accordingly this approach appears to be promising for the characterization of the information flow in brain networks.

Forecasting Brain Activity Based on Models of Spatio-Temporal Brain Dynamics: A Comparison of Graph Neural Network Architectures

“…In this spirit, propagating the information between ROIs based on their SC or structural CE similarity can give us a multimodal perspective of such a directed relationship between brain areas. In the following we choose a perturbation base approach to reconstruct the amount of information one ROI carries about other ROIs [85,79]. By learning a function h(•), the GNN models try to infer from an input sequence of neural activity states [x (1) , .…”

“…Different brain areas communicate via bioelectrical signals transmitted along neuronal axons and collected by neuronal dendrites. Spatio-temporal GNNs provide a novel possibility to incorporate such a structural scaffold into a graph-based prediction model [79]. Due to cognitive information processing in the brain, the spatial interactions of the activity distribution changes dynamically.…”

“…Spatio-temporal GNNs thus encompasses both the information about the layout of the physical scaffold encoded by the graph structure and the dynamical information about temporal activity correlations. Recently we used a DCRNN architecture to model the spatio-temporal brain dynamics in resting state fMRI [79]. In this study, spatial dependencies of brain activities were modeled via diffusion convolution operations based on the anatomical connectivity and the temporal dynamics of the graph signal were captured in an RNN based model architecture [46].…”

Efficient graph convolutional networks for seizure prediction using scalp EEG

“…Specifically, the flatness of Euclidean spaces means that certain operations require many dimensions and complex computations to perform, whereas non-Euclidean spaces may perform these operations more flexibly with fewer dimensions. The graph structure breaks the uniform distribution of the commonly used Euclidean grid structure and can better represent the structural connections and functional realization of brains (Wein et al, 2021 ). It has been used in the diagnosis of autism spectrum disorders (Yang et al, 2021a ), fMRI analysis (Li et al, 2019b , 2021b ), brain network analysis (Wu et al, 2021 ; Royer et al, 2022 ), brain-computer interface decoding (Feng et al, 2021 ; Che et al, 2022 ), emotion classification (Liu et al, 2022 ), and epilepsy detection (Zhao et al, 2021b ).…”

Learning Dynamic Graph Representation of Brain Connectome with Spatio-Temporal Attention

“…Learning the representation of the brain connectome has both positive and negative potential social implications, as it can be linked to the search for new biomarkers of a particular phenotype. Accordingly, the current trend in studies attempting to apply GNN to the brain connectome is to input the FC graph from either resting-state [19,20,2,24,18,35,36] or task fMRI data [21][22][23] and predict a particular phenotype of the subjects, such as gender [20,2,18,17] or presence of a specific disease [20,24,[21][22][23]17]. While these studies have shown potential strengths and opportunities for learning the network representation of the brain, they also suggest limitations of current GNN-based methods.…”