We apply convolutional neural networks (ConvNets) to the task ofdistinguishing pathological from normal EEG recordings in the Temple UniversityHospital EEG Abnormal Corpus. We use two basic, shallow and deep ConvNetarchitectures recently shown to decode task-related information from EEG atleast as well as established algorithms designed for this purpose. In decodingEEG pathology, both ConvNets reached substantially better accuracies (about 6%better, ~85% vs. ~79%) than the only published result for this dataset, andwere still better when using only 1 minute of each recording for training andonly six seconds of each recording for testing. We used automated methods tooptimize architectural hyperparameters and found intriguingly different ConvNetarchitectures, e.g., with max pooling as the only nonlinearity. Visualizationsof the ConvNet decoding behavior showed that they used spectral power changesin the delta (0-4 Hz) and theta (4-8 Hz) frequency range, possibly alongsideother features, consistent with expectations derived from spectral analysis ofthe EEG data and from the textual medical reports. Analysis of the textualmedical reports also highlighted the potential for accuracy increases byintegrating contextual information, such as the age of subjects. In summary,the ConvNets and visualization techniques used in this study constitute a nextstep towards clinically useful automated EEG diagnosis and establish a newbaseline for future work on this topic.
展开▼
