Multi-shell connectome DWI-based graph theory measures for the prediction of temporal lobe epilepsy and cognition

Abstract

Temporal lobe epilepsy (TLE) is the most common epilepsy syndrome that empirically represents a network disorder, which makes graph theory (GT) a practical approach to understand it. Multi-shell diffusion-weighted imaging (DWI) was obtained from 89 TLE and 50 controls. GT measures extracted from harmonized DWI matrices were used as factors in a support vector machine (SVM) analysis to discriminate between groups, and in a k-means algorithm to find intrinsic structural phenotypes within TLE. SVM was able to predict group membership (mean accuracy = 0.70, area under the curve (AUC) = 0.747, Brier score (BS) = 0.264) using 10-fold cross-validation. In addition, k-means clustering identified 2 TLE clusters: 1 similar to controls, and 1 dissimilar. Clusters were significantly different in their distribution of cognitive phenotypes, with the Dissimilar cluster containing the majority of TLE with cognitive impairment (χ2 = 6.641, P = 0.036). In addition, cluster membership showed significant correlations between GT measures and clinical variables. Given that SVM classification seemed driven by the Dissimilar cluster, SVM analysis was repeated to classify Dissimilar versus Similar + Controls with a mean accuracy of 0.91 (AUC = 0.957, BS = 0.189). Altogether, the pattern of results shows that GT measures based on connectome DWI could be significant factors in the search for clinical and neurobehavioral biomarkers in TLE.

Keywords: SVM; TLE; connectome DWI; graph theory; k-means phenotypes.

Fig. 1

(left) GT measures across the Similar cluster (vertical stripes), the Dissimilar cluster (horizontal stripes), and controls (no pattern); and (right) plot representing the clusters separation.

Fig. 2

Community structure of thresholded msDWI-based matrices in (left) controls, (middle) Similar cluster, and (right) Dissimilar cluster. Nodes with the same grayscale tone belong to the same module.

Fig. 3

Proportion of subjects from neuropsychological clusters in the connectome DWI clusters. Generalized cognitive impairment (CI) in stripes, focal CI in dots, no CI in squares.

Fig. 4

Illustration of correct classification (CC) by the SVM algorithm by cluster group: controls (top), Similar cluster (middle), and Dissimilar cluster (bottom).

Fig. 5

ROC curve for 10-fold CV SVM analysis to predict (left) TLE versus control participants and (right) Dissimilar versus Similar combined with control participants.