Addressing Inter-Patient Variability in EEG: Diversity-Enhanced Data Augmentation and Few-Shot Learning-based Epilepsy Detection

Electroencephalogram (EEG) signals are a key source in epileptic seizure recognition. The patient-specific data scantiness in EEG signals, referring to the inter-patient variability of EEG, hinders the accurate recognition of epileptic seizure patterns. This work presents a multi-class epileptic seizure detection scheme with a two-step solution for addressing the data scantiness and inter-patient variability constraint. The proposed Diversity-enhanced data Augmentation and graph theory-assisted FEw-shot Learning for Multi-class seizure detection (DAFEM) approach incorporates diversified data augmentation, graph theory-based feature extraction, and few-shot learning-based multi-class classification. Initially, the data augmentation utilized a Generative Adversarial Network (GAN) with diversified EEG sample generation to conquer EEG data scarcity. Subsequently, it extracts the potential set of features from the augmented data using the graph theory method based on the analysis of inherent dynamic characteristics of EEG data. In particular, to recognize the marginal and the drastic temporal fluctuations in EEG data patterns, it performs the Temporal Weight Fluctuation (TWF) in addition to the feature extraction scores. The data scarcity in the epileptic seizure classes is handled by adopting the few-shot learning strategy. By modeling the Siamese neural network for the multi-class classification of epilepsy, it discriminates the normal, preictal, and ictal patient samples over the constraint of inter-patient variability of EEG data. Finally, the proposed work is tested with two authoritative EEG datasets. The experimental outcomes illustrate that the proposed DAFEM yields 2.73% and 4.5% higher recall on Bonn and CHB-MIT datasets, respectively.