Authors :
Presenting Author: Ryan Bose-Roy, BS – Yale University
Hitten Zaveri, PhD – Yale School of Medicine, New Haven, CT, USA
Rationale:
Localizing seizure onset areas and mapping seizure spread in epileptogenic brain networks is a challenging and intense area of research. Detailed quantitative mapping of these complex, nonlinear spatiotemporal interactions between brain regions through data-driven approaches may offer measures of brain connectivity that can identify and distinguish neural activity patterns between epilepsy patients and non-epilepsy patients, and potentially help elucidate areas of seizure onset. While resting state functional Magnetic Resonance Imaging (rsfMRI) offers the potential for complete mapping of the brain with high spatial resolution, traditional linear models may fail to capture the complex relationships within fMRI. The advent of new transformer and graph neural network-based models, combined with graph theoretic properties such as maximum flow and graph entropy, may allow for new methods of leveraging spatiotemporal data to identify seizure onset areas.
Methods:
Connectivities of epilepsy and control patients were generated from fMRI resting scans parcellated into the Yale Brain Atlas (YBA), a 690 parcel atlas. Functional connectivity matrices were obtained through pairwise correlations across parcels, and effective connectivity matrices were determined through windowed bivariate transfer entropy. Blood Oxygen Level Dependent (BOLD) responses of the 690 Yale Brain Atlas parcels over time in each of 30 epilepsy and 28 control subjects were forecasted using a Hopfield Transformer model. A separate model was trained on each subject, using the past 34 timesteps to forecast the next 3 timesteps across all parcels. After model training, the importances of each parcel to the activity of other parcels were determined by perturbing the activity of a single parcel in the input and observing the change in the trained model’s predictions. Various graph theoretic measures, including eigencentrality, maximum flow, and graph entropy, were computed as a means of assessing and comparing the capacities of pFCs and cFCs to differentiate between epilepsy and non-epilepsy subjects.
Results:
Hopfield attention transformers achieved superior forecasting accuracy compared to Vanilla attention models. The connectivity matrices achieved from both transformer-based approaches exhibited remarkable anatomical correspondence, with K-means clustering consistently segmenting networks into 34 clusters matching anatomical regions in the YBA. Graph-theoretic analyses successfully differentiated epilepsy patients from controls using entropy (92% accuracy), maximum flow (83% accuracy), and eigenvector centrality (73% accuracy). Epilepsy patients showed significantly different regional network properties, particularly in the left inferior temporal-occipital region, temporal pole, and temporal body. The framework also demonstrated potential for identifying seizure onset regions.Conclusions:
This transformer-based perturbation approach offers a novel methodology for inferring causal relationships in functional brain networks that captures anatomical organization and pathological alterations, with significant potential for advancing brain connectivity understanding and developing rsfMRI biomarkers for epilepsy.Funding: None.