Differential quadruple pattern: A new EEG signal classification framework

Abstract

EEG signals are the letters of the brain and reflect neural activity. Abnormal EEG patterns indicate brain disorders such as epilepsy. Recently, machine learning has enabled automated EEG interpretation with high accuracy. This study introduces an explainable EEG classification model based on feature engineering. A novel feature extractor, Differential Quadruple Pattern (DiffQuadPat), is proposed. DiffQuadPat computes relations between four channel values using difference-based transformations and combinational transition tables. Feature selection is performed by Cumulative Weight Neighborhood Component Analysis (CWNCA), and classification is achieved with t-algorithm-based k-Nearest Neighbors (tkNN). For interpretability, Directed Lobish (DLOB) is used to produce symbolic explanations. The proposed DiffQuadPat-centric XFE framework was validated on two EEG datasets: Amyotrophic Lateral Sclerosis (ALS) and neonatal epilepsy detection. The model achieved over 98% accuracy under 10-fold cross-validation. Furthermore, cortical and hemispheric connectome diagrams were generated, enabling transparent visualization of brain-level interactions.