Abstract
Epilepsy is a neurological disease that affects around 50 million people of all ages worldwide. In this study, five deep learning networks were compared to determine the best performance in seizure detection using electroencephalogram raw signals from the TUH EEG Seizure Corpus database. The methodology included three strategies for reducing the high computational cost of training a time series data: Extracting epileptic features from patient signals by concatenating all seizure events into a shorter single one, selecting signals of duration greater than 180 seconds, and generating two randomized groups based on patient and non-patient (control) signals for larger and shorter supervised training-validation processes. Finally, the models were evaluated using two groups, one is formed of patient-control data and the other using only patient data. The results showed that a simple LSTM-based network, a hybrid one and the reported ChronoNet achieved the best metric performance for a binary classification, with up to 71.50 % of sensibility and 83.70 % of specificity for a patient-control detection; and up to 56.60 % of sensibility and 95.90 % of specificity for a patient-specific detection. In conclusion, deep learning-based models might automate seizure detection in order to improve epilepsy diagnosis and accelerate early treatments using electroencephalogram signals.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have