Abstracts

Rationale: The epileptogenic alterations in brain networks may facilitate seizure spread through connections between seizure onset areas and other brain regions. Localizing these areas may prove useful in identifying potential targets for intervention. Although resting state functional Magnetic Resonance Imaging (fMRI) has indicated alterations in functional connectivity at seizure onset areas compared to other areas of the brain, identifying these areas is challenging due to the complex spatiotemporal dependencies between brain regions. Deep learning has enabled performance leaps in the analysis of high-dimensional fMRI, but the advent of new transformer and graph neural network-based models allow for the exploration of new methods of leveraging spatiotemporal data for node classification to identify seizure onset areas.

Methods: Connectivity matrices of epilepsy and control patients were generated from parcellated fMRI resting scans. Cross-correlations provided a measure of functional connectivity while bivariate transfer entropy provided a non-linear measure of effective connectivity from this data. Time-varying graphs were also generated by grouping the parcellated fMRI data into sets of 20 timesteps and calculating the connectivities for each group. The seizure onset areas of these patients were obtained from patient clinical histories and mapped to regions in the 690-parcel Yale Brain Atlas. A graph convolutional neural network was used to classify seizure onset areas and non-seizure onset areas. Other temporal graph neural network models, including a Graph Convolutional-Long Short Term Memory (GCN-LSTM) hybrid and a Graph Convolutional-Hopfield Attention network (GCN-H) were also utilized for node classification.

Results: A graph convolutional neural network model was run on resting state fMRI from 30 temporal lobe epilepsy and 30 neurotypical patients with max-min normalized activity across at least 200 timesteps, with edges generated from cross-correlation or transfer entropy. One concern with the node classification task was that there are far fewer seizure-onset areas compared to the 690 parcels in the brain atlas. A graph-specific self-supervised edge augmentation method based on prior literature was also employed to help the model deal with class imbalances across the 690 nodes, and the resulting accuracy of this model was 83%. A GCN-LSTM and a GCN-H, each running on a single CPU, was able to generate edge and node predictions for the graphs over time with cross-entropy losses of 0.04 and an accuracy of 76%. Structural connectivity information obtained from the Human Connectome Projects was also incorporated as an embedded feature vector for edges in the graph neural networks.


Conclusions: This project uses graph convolutional networks and temporal graph convolutional networks to model resting state fMRI activity connectivity matrices and predict seizure onset areas with high accuracy. These findings are significant because identifying these areas may offer insights into potential treatment avenues as well as novel patterns of seizure activity.


Funding: None.