Classification of Noisy Epileptic EEG Signals Using Fortified Long Short-term Memory Network

Recent work suggests that machine and deep learning models are prone to EEG artifacts and have staggering performance drops when used to classify EEG signals rich of noise. This particularly affects real-time performance of EEG monitoring systems such as brain-computer interfaces, thus rendering their applications in uncontrolled environments useless. These limitations have motivated efforts to develop fortification layers that leverage manifold learning in the lower dimensions to possibly improve the performance and the robustness of any deep learning model by separating off-manifold data points from the dense probability mass. The present study aimed to show that the fortification layer can learn the latent structure of an EEG dataset and that this can help increase the robustness of the classifier when tested on the same dataset contaminated with varying noise. In order to evaluate the performance of the proposed model, different artifacts were synthesized with low bandpass filters to mimic biological and Gaussian white additive noise. Results showed that the EEG signals used in this study followed the manifold assumption, and that the fortification layers learnt the lower discriminative structure from the raw denoised EEG signals. However, this did not significantly increase the robustness of the model to the noise.