Incorporating Hand-crafted Features to Deep Neural Networks for Seizure Prediction

Brain computer interface (BCI) provides effective communication between the brain and a machine. BCI-aided systems have been used for epilepsy control by performing seizure prediction and feedback treatment in a closed-loop approach, in which seizure prediction plays an important role. However, automatic seizure prediction still faces difficulties. Most existing seizure prediction approaches use empirical features which are effective and interpretable, however, they are hard to compose. Recent deep neural network-based methods can learn effective features in a data-driven way, but they are usually not sufficiently trained given limited training data. In this paper, we aim to construct effective seizure prediction methods by combining the strengths of both empirical features and deep models. We construct different deep architectures to incorporate empirical features to different layers in the deep neural networks. Experiments are carried out on nine patients of the Freiburg dataset and achieve a best average F1 score of 0.8622 which outperforms empirical features-based methods and deep neural networks. The results also indicate that incorporating the empirical features in the front layers of the networks can obtain better seizure prediction performance.