Abstracts

Rationale: Resective brain surgery is the most effective therapy for controlling seizures in patients with drug-resistant focal epilepsy. Accurate localization of the epileptogenic zone (EZ) is crucial for improving outcomes following resective surgery. The gold standard for localizing the EZ involves analyzing seizure onset and propagation patterns recorded through stereotactically implanted invasive electrodes, a process known as stereo-electroencephalography (SEEG). Currently, the localization of the EZ using SEEG recordings relies on the manual, qualitative, and subjective interpretation of complex signal patterns associated with seizure onset and evolution. Visual SEEG interpretation requires substantial investment in resources and clinical expertise, thereby reducing the generalizability of these approaches to centers where such expertise is not available. To address these challenges, we propose an explainable convolutional neural network (CNN) architecture that can identify SEEG contacts sampling the EZ.


Methods: We studied 23 patients with temporal lobe epilepsy between 2016 and 2017 who underwent resective surgery following SEEG evaluation and were seizure-free for at least one year. Eighteen of the 23 patients were used for training, and five patients were left out as unseen data to test the trained model. We identified SEEG contacts sampling the EZ and the non-epileptogenic zone for all patients in the training dataset. To address class imbalance, as there were fewer EZ contacts, we oversampled the EZ contacts.

We constructed a clinically grounded deep learning CNN architecture tuned to detect epileptic abnormalities, the presence of oscillatory activities, and their sequential ordering and evolution. The deeper layers of the architecture include self-attention layers inspired by the transformer model. The performance of the trained classifier was assessed by estimating the area under the receiver operating curve (AU-ROC) and accuracy on unseen test data. Finally, we used GradCam to visualize the important SEEG patterns that the model deemed critical for classification.


Results: After aggregating data across seizures, we obtained approximately 15,000 ictal patterns from contacts sampling both the EZ and the non-epileptogenic zone. The CNN model achieved an accuracy of 91% and an AU-ROC of 0.88 on unseen test data (Figure 1). We found that critical features associated with the positive prediction of EZ contacts were related to the presence of preictal spiking and fast activity during ictal onset and evolution (Figure 2A-B). Additionally, we observed subtle fast activity during the preictal phase, which was also deemed important by the model for classifying EZ contacts (Figure 2C).


Conclusions: Our CNN architecture design successfully classifies SEEG contacts sampling the EZ with a low margin of error. The accuracy of this model can be improved by increasing the data size, patient population, and seizure patterns. Furthermore, the developed architecture can be easily integrated into clinical workflows, enabling easier interpretation of SEEG recordings and thereby improving efficiency.


Funding: Cleveland Clinic Neurological Institute Transformative Neuroscience Development Program