Abstracts

Rationale:
Despite its name, focal epilepsy is known to affect large scale functional networks in the brain. Better characterization of these changes can provide new tools for disease classification and biomarkers. Functional gradients are a lower dimension representation of functional networks that can be used to decrease the complexity of whole-brain functional connectivity analysis. Here we investigate the changes in large scale brain networks represented by functional gradients in adult and pediatric patients with focal epilepsy. We use supervised machine learning models to distinguish patients from control.
Method:
Resting state functional MRI data from 64 adult and 21 pediatric patients with refractory focal epilepsy (74% with temporal lobe epilepsy), and 19 adult and 15 pediatric control subjects from 3 epilepsy centers were included. Some subjects contributed more than 1 fMRI run to the dataset. The fMRI data was obtained with variable acquisition details over the years and in different 3T scanners, however was all processed in the same pipeline to produce comparable functional connectivity matrices, defined by Schaefer functional cortical atlases. Preprocessing and functional connectivity analysis were performed using validated pipelines (fmriprep and xcpEngine, respectively). Gradients were calculated in Matlab (R2019a) using the BrainSpace toolbox. The top 5% connectivity values were used to extract the first 10 gradients. The gradients were parcelated into 100, 200, or 400 regions and each parcel value was considered a feature (1000-4000 features per fMRI run). A linear support vector machine classifier (SVM) was trained using randomly sampled 60% of the dataset, while the other 40% were used for testing the model and reporting accuracy. We also aimed to determine seizure laterality using the same method among epilepsy patients.
Results:
The 85 patients and 34 controls contributed to 101 and 96 runs of fMRI data, respectively. Using the first 10 gradients for each run, SVM for epilepsy vs control classification reached training accuracies of up to 77%.  When applied to independent sample classification accuracies of 65, 73, and 69% were achieved for 100, 200, and 400 parcels, respectively. Within the epilepsy group, SVM did not achieve accuracies beyond chance levels.
Conclusion:
Our findings demonstrate that major functional gradients are mostly preserved in epilepsy, however some functional characteristic differences are captured by functional gradients. These can be used to distinguish epilepsy patients from controls in heterogeneous datasets. Lateralization models were limited partially due to small number of right onset patients. Further understanding of changes in functional gradients in epilepsy and the use of larger datasets may improve classification accuracy.
Funding:
:NIH NCATS UL1TR001876 / KL2TR001877 award to TG