Abstracts

Rationale: The expert epileptologists detect seizures directly by visually analyzing EEG plot images, unlike the automated methods that analyze spectro-temporal features or complex, non-stationary features in EEG signals. So, seizure detection could benefit from convolutional neural networks (CNN) as their performance in visual recognition will be comparable to that of humans. To reduce false alarm rate, non-seizure abnormalities like the physiological artifacts should be included in the training data. To automatically find non-seizure abnormalities, we attempted to use autoencoder (AE), which compresses the data with minimum data loss. We demonstrated that false alarm rates were improved by adding non-seizure abnormalities with high AE errors in training data of CNN. Methods: The data, 1124 hours data from 24 patient with epilepsy include 97 seizures, after preprocessing, was segmented by one-second time windows and each segment was converted to a vector. The autoencoder trained with the vectors and AE error used for finding abnormalities. We used an image-based seizure detection by applying CNN to long-term EEG that included epileptic seizures. After filtering, EEG data was divided into one-second segments and converted into plot EEG images. We made 2 training dataset, in the first one the non-seizure data was randomly selected from inter-ictal. In the second one, the half of non-seizure data selected from non-seizure abnormalities classified by autoencoder and the rest selected randomly from the remaining inter-ictal data. The CNN trained separately with these 2 training datasets, and the result of seizure detection was compared. Results: The median of detected seizure rate by minutes was 100% for both training datasets. AE improved the false alarm rate: The median of false alarm rate was 0.13 false alarm per hours with AE, while 0.17 false alarm per hour without AE. Conclusions: The artificial visual recognition by CNN was a powerful tool for seizure detection, that currently relies on skillful visual inspection of expert epileptologists in clinical diagnosis. The performance was improved by adding non-seizure abnormalities with high AE errors in training data of CNN. Funding: This work was partially supported by KAKENHI grants (26242040, 17H04305) and Tateisi Science and Technology Foundation.