OPTIMIZATION OF THE DEEP NEURAL NETWORKS FOR SEIZURE DETECTION

The goal of the present study was to optimize model selection and data preparation procedures for seizure detection in patients with epilepsy on wearable EEG data for the “ICASSP Signal Processing Grand Challenge’. We used the SeizeIT1 dataset for machine learning model training and the SeizeIT2 validation dataset provided by the organizers of the Challenge. We tested more than 100 machine learning architectures and hyperparameter combinations to achieve the most accurate, robust, and generalizable performance in seizure detection tasks. The best models included the spectral transformation of raw EEG data for the DCNN model input, using correct crossvalidation procedures, tuning data sampling for class imbalance problems, and data augmentation procedures. The best model’s sensitivity and normalized false alarm (FA/h) rate were 77.78% and 3.35 respectively for the STFT-based DCNN developed by the authors. The best sensitivity of the adapted ChronoNet model chosen for testing the robustness of the models was 80% with FA/h = 9.45.