Abstracts

Rationale: Children with drug-resistant focal epilepsy (DRE) often suffer from highly frequent seizures that can interfere with the development of white matter connections, which could be effectively characterized by measuring lateralization index (LI) of diffusion weighted imaging connectome (DWIC). This study evaluates the feasibility of our novel deep learning technique to predict clinically observed seizure frequency (SF) values (i.e., total number of seizures per month) of individual DRE children using their DWIC-measured LI values.

Methods: Fifty DRE children (age: 11.7±3.5 years; 24 boys; 26 left and 24 right hemispheric epilepsy) and 29 healthy children (HC, age: 11.6±3.3 years; 14 boys) had 3T whole brain DWI tractography scans. Left and right intra-hemispheric networks, G=(N,S), were constructed, where N_{i or j=1-45} is a set of 45 nodes representing 45 homologous regions and S is an adjacency matrix with elements S_ij, representing average fractional anisotropy values of pair-wise connection between N_i and N_j. At each N_i, current flow betweenness (CFB) was measured as a marker assessing both number and strength of neighboring connections. The relationship between SF and CFB_LI (i.e., [non-epileptic side–epileptic side]/[non-epileptic side+epileptic side]) was evaluated using Pearson’s correlation. Z-scores of CFB_LI, indicating deviation from those of age-matched HC, were compared between two groups (SF < 30 vs. ≥30/month). SF values of individual patients were predicted using a deep residual neural network (DRNN) with CFB values of individual N_i nodes and LI_ij (i.e., LI matrix of S_ij). Briefly, the DRNN was trained using 3-fold cross validation in a model cohort (n=33 patients). The reproducibility of the trained DRNN was confirmed in a validation cohort (n=17 patients) that was not included in the model cohort. The Pearson’s correlation coefficient, R, and mean absolute error (MAE) between predicted and observed SF were used as a measure of the prediction accuracy.

Results: SF was positively correlated with CFB_LI values in insula, hippocampus, superior parietal gyrus, caudate, putamen, thalamus, and middle temporal gyrus (R/p: >0.30/< 0.05). Highly frequent (at least daily) seizures were associated with higher CFB_LI of the insula, hippocampus, superior parietal gyrus, thalamus, and caudate (Fig. 1), suggesting increased CFB in non-epileptic side even compared with age-matched HC. DRNN-predicted SF was significantly correlated with the observed SF in the validation cohort (average R/p:0.70/< 0.01, MAE:22.69). DRNN’s attention layer identified 9 nodes of which LI values were the most predictive of SF (Fig. 2).

Conclusions: This study provides preliminary evidence supporting that highly frequent seizures in the epileptic hemisphere are associated with increased axonal integrity in the homologous regions of non-epileptic hemisphere. We hypothesize that the enhanced lateralization of axonal connectivity could be a potential sign of neural plasticity responding to highly frequent seizures.

Funding: NIH R01NS089659 (to J.J), R01NS064033 (to E.A), and R01NS041922 (to C.J)