Operational Transformer: An investigation of epilepsy detection

Electroencephalography (EEG) signals represent the electrical activities of the brain and have been utilized to assess brain conditions. EEG signals are also crucial for diagnosing epilepsy. However, EEG interpretation is a challenging task. Therefore, new-generation methods should be introduced. The essential goal of this study is to present a new transformer model for multichannel EEG signal classification. A new transformer model has been introduced in this research, termed the Operational Transformer (OpT). To evaluate the classification capability of OpT, a new-generation explainable feature engineering (XFE) framework is presented. The OpT-driven XFE approach comprises four key stages: (i) feature derivation utilizing OpT and a transition table feature extractor to obtain EEG signal attributes, (ii) identification of the most significant features through cumulative weighted iterative neighborhood component analysis (CWINCA), (iii) classification of the selected features via k-nearest neighbors (kNN), and (iv) generation of explainable outputs leveraging the Directed Lobish (DLob)-based interpretation method. These phases were integrated to construct an XFE framework aimed at measuring the efficiency of OpT, which was validated on a publicly available EEG epilepsy dataset. The presented OpT-centric XFE model yielded classification accuracies of 99.99% and 84.74% under 10-fold cross-validation (CV) and leave-one-subject-out (LOSO) CV, respectively. Furthermore, a connectome diagram was generated using DLob for the employed dataset. The computed classification and interpretability results show that the introduced OpT-driven XFE model performs strongly under the reported experimental conditions. The presented XFE model contributes to feature engineering by providing high classification performance and to neuroscience by generating interpretable results utilizing DLob.