PyramidPat explainable feature engineering for multiclass electroencephalography psychiatric disorders: Explainable feature engineering and classification

Abstract

Deep learning has dominated modern machine learning, and feature engineering has often been neglected. Many studies still focus mainly on accuracy. Therefore, explainable artificial intelligence (XAI) methods remain limited. In this research, a new explainable feature engineering (XFE) architecture, named PyramidPat XFE, is introduced. The core component is PyramidPat, a transformation-based feature extractor for multichannel signals. The pipeline has four stages: (1) PyramidPat feature extraction, (2) feature selection with INCA, (3) classification with tkNN (an iterative ensemble kNN), and (4) DLob-based explanation, which converts selected feature identities into lobe-and channel-based sentences. The evaluation is performed on a six-class EEG psychiatric disorder dataset with seven defined test cases. With leave-one-subject-out (LOSO) cross-validation, the proposed model achieves accuracy above 93% in all cases and produces interpretable results for every case. These outcomes indicate that PyramidPat XFE is effective for EEG-based psychiatric disorder classification and for generating compact XAI outputs from EEG signals.