Abstracts

Rationale: West syndrome is an epileptic syndrome with an unfavorable prognosis that usually affects infants under one. If the West syndrome is diagnosed, treatment should be started promptly, and this is because if treatment is delayed, there is a high possibility that sequelae will remain. In order to improve the diagnostic efficiency, we propose a diagnostic assistant method, following the EEG visual diagnosis mechanism of epileptologists that directly processes the plotted EEG image to detect hypsarrhythmia from EEG monitoring data. In addition, it improves clinicians’ confidence in the method by analyzing and interpreting the judgment basis of the machine learning model used, and if the model returns the quantitative result of hypsarrhythmia, the clinicians can evaluate medical treatment efficacy with an EEG after treatment.

Methods: Following the EEG visual diagnosis mechanism of the epileptologists, EEG monitoring data is first plotted as images by a sliding window (ten seconds with one second step), each image has a label of hypsarrhythmia or normal that the clinicians have annotated, and the images can be fed into the convolutional neural network (CNN) model for the classification task, the overall structure of the method is shown in Fig. 1. After getting the model classification results, for each image, we use the Gradient-weighted Class Activation Mapping (Grad-CAM) method for visual analysis of the classification basis. In addition, for each image, the CNN model inferred classification probability is also normalized to a zero to one value that measures the degree of the hypsarrhythmia in the EEG image. A high score means serious hypsarrhythmia.

Results: Interictal epileptic EEG of the West syndrome from six patients are used to evaluate our method (collected from Juntendo University hospital and the data collection was approved by Juntendo University Hospital and Tokyo University of Agriculture and Technology Ethics Committee), and the leave-one-patient-out method was used to evaluate method performance in the cross patient. The classification results are shown in Figure 2A. For the image samples that were classified as hypsarrhythmia, the Grad-CAM method is used to analyze the model classification basis visually, two results are shown in Figure 2B, and the blue color part as the classification basis. Each image also has a classification probability value from the model last layer, and the value was used to quantify the hypsarrhythmia degree in the EEG image, fours results are shown in Figure 2C.

Conclusions: From the experiment results, the EEG image with the CNN model shows an expected performance. Based on the classification results, we use Grad-CAM to analyze the classification basis, and the visualization result can provide more practical reference information to clinicians. The model classification probability is also provided to clinicians as degree quantification that can evaluate medical treatment efficacy.

Funding: This work was supported by JST CREST Grant Number JP-MJCR1784 including AIP challenge program, Japan.