LAUSR

BACKGROUND Early prediction of epilepsy seizures can warn the patients to take precautions and improve their lives significantly. In recent years, deep learning has become increasingly predominant in seizure prediction for its outstanding performance. With the aim of predicting unseen seizures, it is essential to guarantee the generalization ability of the model, especially considering the non-stationary nature of EEG and the scarcity of seizure events in EEG recordings. Stability training against extra perturbations is an intuitive and effective way to improve the model's ability to generalize. Though a great number of deep learning methods have been developed for seizure prediction, their strategies to increase generalization performance focus on improving the model's architecture itself, and few of them pay attention to the stability of the model against small perturbations. NEW METHOD In this study, we propose a novel consistency-based training strategy to address this issue. The proposed strategy underlines that a robust model should maintain consistent results for the same input under extra perturbations. Specifically, during training, we use stochastic augmentations to make the input vary from iteration to iteration and consider the output as a stochastic variable. Then a consistency constraint is constructed to penalize the difference between the current output and previous outputs. In this way, the generalization ability of the model will be fully enhanced. RESULTS To better verify the effectiveness of our proposed strategy, we implement it in two state-of-the-art models with public-available codes, including STFT CNN and Multi-view CNN. Notably, we compare with the first baseline on a scalp EEG dataset and the other on an intracranial EEG dataset. The results show that our strategy could improve the performance significantly for both of them. COMPARISON WITH EXISTING METHODS Our strategy has increased the sensitivity by 7.1% and reduced the false prediction rate by 0.12/h on the first baseline while improving the AUC by 0.020 on the second baseline. CONCLUSIONS This study is easy to implement, providing a new solution to enhance the performance of seizure prediction.