P64. Deep learning for EEG diagnostics

Abstract

Machine learning as a tool for medical diagnostics is gaining increasing interest. For example, a deep convolutional neural network (deep ConvNets) performed at least as well as dermatologists in the diagnosis of different types of skin cancer from images and another deep ConvNet segmented retinal vessels better than human annotators. Therefore, we investigated the potential of deep ConvNets for diagnosis from electroencephalography (EEG) recordings. We used two recently proposed ConvNet architectures that were shown to decode task-related information from EEG at least as well as other established machine learning methods. The ConvNets were trained to distinguish pathological and normal EEG recordings in the Temple University Hospital EEG Abnormal Corpus that contains about 3000 recordings from more than 10 years of clinical practice. The EEG recordings were labelled normal or pathological by a neurologist. The ConvNets classified the recordings more accurately than the only published baseline result on this dataset by a large margin (∼85% vs ∼79% decoding accuracy, resp.). Interestingly, they obtained better accuracies even when only using 1 min per recording in the training process and only 6 s to make the diagnosis. We applied visualizations to explore what the ConvNets learned, finding that they used, possibly together with other features, band power changes in the delta (0–4 Hz) and theta (4–8 Hz) frequency ranges, matching well with the textual medical reports. While the diagnosis of our ConvNets could already be clinically valuable, for example by flagging EEG recordings where there is a disagreement between ConvNet and neurologist for a second analysis, there are promising ideas to further increase their accuracies. For example, using more recently proposed optimization methods such as stochastic gradient descent with restarts could increase the ConvNet performance, while using recurrent neural networks could allow to integrate information on a longer time scale.