A novel feature-to-image decoding framework using NeRV, GAN, and decoder for discriminative epileptic seizure detection from EEG signals

Abstract

Epilepsy is a neurological disorder triggered by abnormal electrical discharges in cortical neurons. This condition can lead to permanent brain damage and irreversible outcomes. In this study, an artificial intelligence-based hybrid model for detecting epileptic seizures using electroencephalography (EEG) signals is proposed. The publicly available Bonn EEG dataset was used for this study. This dataset consists of five EEG categories representing healthy, pre-seizure, and seizure states. In the proposed approach, EEG signals were preprocessed by converting them into image format using the gramian angular field (GAF) method while preserving their temporal structure. GAF-based image clusters were trained using transformer-based models (PiT, PVT) to obtain class-based feature columns. The feature columns obtained from each transformer model were recoded into images using three different neural processing methods (Decoder, GAN, NeRV) to close the semantic gap related to visual interpretability. A discriminative score method was used to select the most representative image for each variant (Decoder-based, GAN-based, NeRV-based) from these images. The selected representative images were classified using a residual-based random forest method to detect epileptic seizure types. In experimental analyses, the selection steps performed to prevent information leakage were limited to training data only. Both holdout and cross-validation strategies were used in the experiment. A general accuracy success rate of 99.81% was achieved with the cross-validation technique, and a general accuracy success rate of 99.76% was achieved with the holdout technique. Although the experimental results are highly distinctive, the dataset used has contributed to the results with a controlled comparison criterion.