Detection of Epileptic Seizures With Autoencoders Working With Multiple EEG Frames

Electroencephalogram (EEG) signals can be used for discriminating between ictal and non-ictal phases, to protect people affected by epilepsy. Recent studies show the ability of autoencoders (AEs) to learn valuable EEG abstractions that can be effectively processed by classifiers. To some extent, information from previous frames could be useful for detection. However, the design should avoid using AEs with large input sizes, to prevent unsuccessful feature learning. This paper discusses AE-based schemes which allow the investigation of multiple successive frames. A hierarchical feature extractor is introduced to correlate long-term information via AEs of reduced input sizes, to support easy training. The experimental results indicate improved detection accuracy for the classification based on the proposed multi-frame hierarchical feature extractor, compared to feature extractors that work on a single frame, or use a priori/ a posteriori aggregation of frames.