Two-stage model for epileptic seizures detection on EEG recordings

The purpose of this study is to analyze the applicability of a two-stage model based on convolutional neural networks to improve the quality of seizure detection on real EEG data. Wavelet analysis is used for time-frequency analysis. To localize epileptic discharges, the seizure detection task was reduced to the classification task where the prediction process consists of two steps: the first model provides coarse predictions which are refined by the second model trained on the first model’s errors. As a result of using the proposed two-stage model, the F1-score metric was improved by about 2% compared to a single coarse model, and at the same time led to a significant increase in false negative predictions, which shows the tradeoff brought by the considered approach.