Towards Explainable Graph Neural Networks for Neurological Evaluation on EEG Signals

TL;DR

A novel approach using Graph Neural Networks (GNNs) to predict stroke severity, as measured by the NIH Stroke Scale (NIHSS), is proposed, providing clinicians with a valuable tool for diagnosis, personalized treatment, and early intervention in neurorehabilitation.

Abstract

After an acute stroke, accurately estimating stroke severity is crucial for healthcare professionals to effectively manage patient's treatment. Graph theory methods have shown that brain connectivity undergoes frequency-dependent reorganization post-stroke, adapting to new conditions. Traditional methods often rely on handcrafted features that may not capture the complexities of clinical phenomena. In this study, we propose a novel approach using Graph Neural Networks (GNNs) to predict stroke severity, as measured by the NIH Stroke Scale (NIHSS). We analyzed electroencephalography (EEG) recordings from 71 patients at the time of hospitalization. For each patient, we generated five graphs weighted by Lagged Linear Coherence (LLC) between signals from distinct Brodmann Areas, covering $δ$ (2-4 Hz), $θ$ (4-8 Hz), $α_1$ (8-10.5 Hz), $α_2$ (10.5-13 Hz), and $β_1$ (13-20 Hz) frequency bands. To emphasize key neurological connections and maintain sparsity, we applied a sparsification process based on structural and functional brain network properties. We then trained a graph attention model to predict the NIHSS. By examining its attention coefficients, our model reveals insights into brain reconfiguration, providing clinicians with a valuable tool for diagnosis, personalized treatment, and early intervention in neurorehabilitation.

Figures (7)

• Figure 1: Full pipeline from EEG data collection to model input: (a) EEG recording setup, illustrating the placement of electrodes on the scalp for data collection Shen_2020. (b) The initial fully connected graph obtained after the EEG's preprocessing and the application of eLORETA, representing all possible connections between 84 Brodmann areas in the brain, constructed from the EEG data. (c) Sparsified graph after the rewiring process, serving as input to the model. Blue edges represent structural connections based on spatial proximity, while green edges are functional connections derived from the top 1% of LLC values.
• Figure 2: Patient A, Stroke side, right - NIHSS, 19: (Left) Brain graph analysis based on small-world metrics, specifically using the weighted clustering coefficient for the nodes and edge betweenness for the edges. (Right) Nodes are visualized using weighted in-degree centrality calculated through attention coefficients, while the edges represent the actual values of the attention coefficients. These are the combined graph built from $\mathsf{G}_{\alpha_1}$, $\mathsf{G}_{\alpha_2}$, $\mathsf{G}_{\beta_1}$. For each pair of Brodmann areas, the edge with the highest value was selected: $e_{ij} = \max(e_{ij}^{\alpha_1}, e_{ij}^{\alpha_2}, e_{ij}^{\beta_1})$.
• Figure 3: Illustration of a multi-layer network with three layers. Each layer is associated to a LLC graph for a specific frequency band. To allow for cross-layer communication, each node is connected to all the nodes with the same label of different layers. To reduce clutter, vertical dotted lines are placed only between successive layers.
• Figure 4: Distribution of NIHSS across three categorized severity groups: : Class A, for NIHSS < 9; Class B, for 9 $\leq$ NIHSS < 16; and Class C, for NIHSS $\geq$ 16
• Figure 5: (Left) Patient A, Stroke side, left - NIHSS, 21. (Right) Patient B, Stroke side, right - NIHSS, 21. In both graphs, the color and size of the nodes are determined by the weighted in-degree centrality. These are the combined graph built from $G_{\alpha_1}$, $G_{\alpha_2}$, $G_{\beta_1}$. For each pair of Brodmann areas, the edge with the highest attention coefficient was selected: $e_{ij} = \max(e_{ij}^{\alpha_1}, e_{ij}^{\alpha_2}, e_{ij}^{\beta_1})$.