LAUSR

Introduction In patients with drug-resistant focal epilepsy, surgery can be considered. The goal is to remove the epileptogenic tissue, while sparing the eloquent cortex. Prior to surgery, a prolonged electroencephalography (ECoG) recording can assist in the delineation of epileptogenic tissue and functionality of the surrounding cortex. During these recordings, Single Pulse Electrical Stimulation (SPES) of the intra-cranial electrodes is performed to evoke pathological responses from the epileptogenic tissue, which occur >100 ms after stimulation. These responses are called delayed responses (DR). In the UMC Utrecht, they are visually analyzed by use of time-frequency images from approximately 2 s. around stimulation. Each image is scored by two human observers on the presence of an evoked DR in three different frequency bands, namely Spikes (S, 10–80 Hz), Ripples (R, 80–250 Hz) and Fast Ripples (F, 250–510 Hz). This visual analysis is very labor intensive. Therefore, we trained a Support Vector Machine (SVM) and a Convolutional Neural Network (CNN) to mimic the human observer in scoring the images. Methods The training data consisted of 47,197 images from 15 patients, with the consensus of two human observers as ground truth. The algorithms were tested on a total of 11,394 images from 4 other patients. For the SVM, 9 features were defined and extracted from each image. The CNN used the whole image as an input. Classification was based on 5 different outputs. F1 scores were calculated for all classes separately. Results The CNN achieved an average F1 score of 0.55. The SVM did slightly better with 0.60. Sensitivity and precision for the DRs were 0.88 and 0.65 for the SVM vs 0.96 and 0.42 for the CNN. Conclusion Two machine learning algorithms were trained to score time-frequency responses of SPES. Both models showed a high sensitivity but a lower specificity for DRs. This was more pronounced for the CNN than for the SVM. Nonetheless, a drastical decrease in time and effort needed for the analysis of SPES is already achieved. More data of the underrepresented classes should be created for the algorithms to improve. The algorithms will be applied to additional patient data to see whether the agreement with human observers is comparable with the inter-rater agreement. Future research will aim at relating the found DRs to the seizure onset zone.