Abstracts

Rationale: Brain structure as captured by diffusion tensor imaging (DTI) can provide relevant pathophysiological information regarding different neurological disorders. Graph theory (GT) methods have helped in the characterization of global, regional, and topological brain properties on structural imaging making possible the understanding of the effects of such disorders. Further, machine learning (ML) algorithms have been useful in identifying intrinsic within-group differences in temporal lobe epilepsy (TLE) regarding cognition, morphology and function, which helped identify different levels or phenotypes within TLE patients. In this study, we performed ML analyses on GT metrics extracted from DTI data from TLE patients in order to investigate how sensitive the combination of ML on GT measures is regarding intrinsic group subdivision in TLE, and distinction between TLE and controls.

Methods: Unsupervised K-means clustering was applied on brain structural data from TLE patients using GT metrics as features in order to identify intrinsic network phenotypes and characterize their clinical significance. Moreover, Support Vector Machine (SVM) algorithms were applied to the same GT measures in order to investigate if group membership (TLE or control) could be predicted. Participants were 89 TLE and 50 healthy controls from the Epilepsy Connectome Project (ECP), and harmonized DTI symmetric matrices of 164 nodes were calculated for each participant. Global efficiency, transitivity, modularity index and a combined centrality measure (i.e., an average of degree and betweenness centrality) were the GT measures used on the K-means clustering algorithm and on the SVM analysis. _x000D_
Results: K-means clustering rendered 2 clusters on the TLE group: one comparable to controls (i.e., Similar cluster) and one deviating from controls (i.e., Dissimilar cluster) (Figure 1). Clusters were significantly different (=6.641, p=0.036) in terms of the proportion of previously obtained neuropsychological clusters (Hermann et al., 2020). In that case, the Dissimilar cluster contained the majority of TLE participants from the Severeneuropsychological cluster (67%), and the Similar cluster included the majority of the Unimpairedneuropsychological cluster (77%). In addition, SVM was able to predict group membership with an accuracy of 0.69 with an area under the curve (AUC) of 0.9091 and a Brier score of 0.0909._x000D_
Conclusions: GT measures extracted from DTI matrices were able to identify 2 structural phenotypes in TLE participants. Such clusters were significantly different in terms of previously obtained neuropsychological phenotypes, therefore, they seem to capture cognitive discrepancies in TLE. Furthermore, SVM was able to discriminate between groups therefore capturing intrinsic structural differences between controls and TLE participants.

Funding: Research reported in this publication was supported by the National Institute of Neurological Disorders and Stroke and the National Institute of Aging of the National Institutes of Health under award numbers R01-1NS111022, R01NS117568, R01NS123378, R01NS105646, R01AG063849, and by U01NS093650 (Epilepsy Connectome Project).