Abstracts

Convolutional Neural Networks for Detecting Epileptiform Discharges in Scalp EEG

Abstract number : 3.144
Submission category : 3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year : 2016
Submission ID : 197621
Source : www.aesnet.org
Presentation date : 12/5/2016 12:00:00 AM
Published date : Nov 21, 2016, 18:00 PM

Authors :
Michel J.A.M. van Putten, University of Twente and Medisch Spectrum Twente, Netherlands

Rationale: Clinical use of computer assisted detection of epileptiform discharges is limited, and visual assessment still outperforms current algorithms. Similar to the success of recognizing handwritten digits and classification of images, I propose to use deep learning algorithms for the detection of epileptiform discharges. Such an approach allows training with a very large number of multi-channel EEG examples where the spatial structure can be preserved using convolutional neural networks (CNN). Methods: Two-second 19-channel epochs of clinical EEG recordings sampled at 256 Hz, all available in our digital database were labeled as one of five categories: normal , presence of epileptiform discharges, EMG artifact, diffuse slowing, or eye blinks, using a set of 12 routine EEGs. EEG snippets of 2 s were labeled and divided into a training (n=383) and evaluation (n=164) set. A CNN was implemented in python using Keras on top of Theano. The CNN was run on an iMac using a CUDA-enabled NVIDIA GPU. Results: Training the CNN took approximately 80 seconds. The classification error was approximately 15%, where most errors resulted in erroneous labeling of EMG artifacts as a normal background pattern or epileptiform discharges. An example is presented in Figure 1. Conclusions: Convolutional neural networks preserve the spatial characteristics of input data, making them very suitable for processing multichannel EEG. In this pilot study, epileptiform discharges were satisfactory classified after training the network with a relatively small number of examples . At present, we are extending our dataset to create over 10-thousand labeled EEG epochs to be used for training, including digitized EEG examples from textbooks. As a next step, unsupervised learning may be implemented. Our final goal is to obtain sensitivities of 98% or better, with a false detection frequency of less than 1 per hour of EEG recording. Funding: A continuation of this project will be funded by the Dutch Epilepsy Foundation
Neurophysiology