Self-Organized Explainable Symbolic EEG Feature Engineering Using Directed Lobish and Previous-Next Pattern: A Proof-of-Concept Study on Dementia Classification

Objective: Dementia is a prevalent neurological disorder, and accurate detection is essential for timely intervention. This study aims to design an interpretable symbolic feature engineering framework for EEG-based dementia classification and to examine how its performance depends on the validation protocol. Methods: We used a publicly available EEG dementia dataset including Alzheimer’s disease (AD), frontotemporal dementia (FD), and healthy controls. A self-organized explainable feature engineering (SOXFE) model was introduced with five phases: (i) feature extraction using Bipartite Channel Transformer (BCT) and Previous-Next Pattern (PNPat), (ii) feature selection via cumulative weight-based iterative neighborhood component analysis (CWINCA), (iii) explainable output generation using Directed Lobish (DLob), (iv) classification through an ensemble tkNN classifier, and (v) information fusion with combinational majority voting (CMV). Performance was evaluated with segment-wise 10-fold cross-validation, leave-one-subject-out (LOSO) cross-validation, and nested participant-grouped 10-fold cross-validation. Results: Segment-wise 10-fold cross-validation, which allowed segments of one participant in both training and test folds, yielded accuracies of 99.12–99.74%. Participant-independent accuracy was substantially lower: 51.19–63.81% for LOSO and 49.00–62.50% for nested participant-grouped validation at the segment level. Participant-level LOSO accuracy equaled the majority-class rate for dementia versus healthy controls (67.05%) and fell below it for AD versus FD (57.63% versus 61.02%). No exact McNemar comparison showed an accuracy advantage over the majority-class baseline. DLob outputs were produced as feature-level visualizations. Conclusions: The findings support the feasibility of interpretable, algorithm-derived symbolic feature representations. They do not establish reproducible disease-related EEG characteristics or clinical diagnostic utility. DLob outputs describe the channel transitions used by the model. Independent cohorts and stronger participant-independent performance are required before clinical application.