Authors :
Presenting Author: Elakkat D Gireesh, MD – AdventHealth Orlando
James Baumgartner, MD – AdventHealth Orlando; Po-Ching Chen, PhD – AdventHealth Orlando; Varadraj Gurupur, PhD – University of Central Florida; Ki Hyeong Lee, MD – AdventHealth Orlando; Joohee Seo, MD – AdventHealth Orlando; Holly Skinner, DO – AdventHealth Orlando
Rationale:
Recent advances in the robotics and imaging technology have enabled precise and extensive stereo EEG (SEEG) electrode placement for intracranial EEG monitoring, as part of presurgical evaluation, which has become an extremely helpful strategy to identify epileptogenic zones. This has paved the way for generation of enormous amounts of data that needs to be analyzed before determination of the epileptogenic focus which can be challenging with naked eye. Appropriately designed deep neural network based tools may be used in addressing this challenge.
Methods:
In this study we report the tools developed in identifying the epileptogenic focus from SEEG data using deep neural network. We used retrospective data from 10 patients who underwent successful epilepsy surgery (seizure outcome: ILAE 1-3), after SEEG monitoring. The interictal EEG data (1 minute of consecutive data, acquired from all the electrodes implantaed) from the initial period of SEEG monitoring was used to develop a multilayer model to predict the seizure onset zones. The higher frequency components of the SEEG (60-600 Hz) was used for developing the model. The class of data (epileptic vs non-epileptic) was decided by the epileptogenic zones as recorded in the clinical records. Deep neural networks (DNN) designed using Python programming language (tensorflow package) in two methods (dense layers, convolutional neural network), were trained with the SEEG data. The training models were initially validated with the patient data used for training. Also, the model was later validated with the SEEG data from seven additional patients, that were not used in the training.
Results:
Testing the model revealed >90 % accuracy in predicting the channels which are potentially epileptogenic. The model was analyzed further to evaluate for understanding the features of SEEG that the DNNs were using to arrive at decisions. We used the Gradient class activation mapping (Grad-CAM) to generate heatmaps, which point to the features of the signal that were relevant for DNN’s decision making. The heatmaps were correlated to the raw signal and the analytical signal calculated from Hilbert transform of the signal. Maximum correlation of heatmap with the processed signal would suggest that the features highlighted in that kind of processing are significantly contributing to the decision making by DNN.
Conclusions:
The model predictions showed string concordance with the conclusion of the clinical teams on the epileptogenic zones. This suggests the potential use of deep learning network models in predicting epileptogenic zones based on interictal EEG data.
Funding: None