Using single pulse electric stimulation and deep learning methods to identify the seizure onset zone
TL;DR
A hybrid convolutional–transformer neural network that classifies whether individual recording contacts lie within the seizure onset zone using the first 100 ms of stimulation-evoked responses indicates that stimulation-evoked responses can contribute to seizure onset zone localization without requiring spontaneous seizure capture.
Abstract
Drug-resistant focal epilepsy affects approximately one-third of epilepsy patients and often requires surgical resection of the seizure onset zone for seizure control. Localizing the seizure onset zone currently depends on prolonged intracranial monitoring and spontaneous seizure capture, a process that is time-consuming and may produce inconclusive results. Prior work has shown that cortico-cortical evoked potentials elicited by single pulse electrical stimulation through implanted stereoelectroencephalography electrodes differ between epileptogenic and non-epileptogenic tissue, but existing computational approaches have been limited to single-center temporal lobe epilepsy cohorts. This thesis addresses that gap by developing a hybrid convolutional–transformer neural network that classifies whether individual recording contacts lie within the seizure onset zone using the first 100 ms of stimulation-evoked responses. The model was trained and evaluated on data from 16 patients across two independent centers, the University of Utah and the University of Missouri. Fifteen patients had clinically defined temporal SOZ, three of whom additionally had extra-temporal SOZ involvement; one patient was non-localized. A total of 65,220 stimulation-response contact pairs were analyzed with a class imbalance ratio of 8.4:1. Using 4-fold repeated stratified group k-fold cross-validation with patient-level splits, the model achieved a mean sensitivity of 0.669, specificity of 0.789, and area under the receiver operating characteristic curve of 0.783 across 20 held-out test evaluations. Expected Gradients attribution analysis identified the 10–50 ms post-stimulation window, corresponding to the N1 evoked potential component, as the primary driver of classification across both temporal and extratemporal regions. These results indicate that stimulation-evoked responses can contribute to seizure onset zone localization without requiring spontaneous seizure capture.