Research Associate Tufts Univeristy Boston, Massachusetts
Rationale: The detection of electrographic seizures using manual inspection of electroencephalography (EEG) recordings is labor intensive and can result in large variability across researchers and clinicians. Accumulating evidence illustrates that the use of deep learning networks can be a promising avenue for automating seizure detection. Here we tested deep learning models on seizure detection accuracy and investigated how they can be combined with manual inspection to improve the efficiency of accurate seizure detection. Methods: The Keras Python library (Tensorflow backend) was used to build deep learning models based on convolutional neural network (CNN) architectures for classifying LFP/EEG segments as seizure or no seizure. The models were trained and tested on LFP/EEG recordings obtained from chronically epileleptic mice that were generated using intra-hippocampal kainate injections. The models were trained on 5 second windows of LFP/EEG data down-sampled (decimated) at 100 Hz on a subset of the dataset. Results: We found that the use of deep neural networks detected seizures with high sensitivity. When paired with manual inspection of detected events, classification efficiency vastly improved ( > 1000x faster) when compared with manual inspection whilst maintaining high-sensitivity and accuracy. Conclusions: Semi-automation of seizure detection using deep learning models vastly reduced manual labor while maintaining a high accuracy of seizure detection. Application of automated deep learning models will benefit the epilepsy community by minimizing bias and improving seizure detection efficiency. Funding: Please list any funding that was received in support of this abstract.: NIH-NINDS
