Compact relative-ordering models for interpretable EEG-based dementia classification

Most EEG dementia classifiers rely on absolute feature magnitudes. We assessed whether pairwise relative-ordering methods, widely used in omics, support compact, interpretable EEG classifiers. We analysed eyes-closed (EC) and eyes-open with photic stimulation (EO/photo) recordings from the same Alzheimer’s disease (AD), frontotemporal dementia (FTD), and cognitively normal (CN) OpenNeuro cohort (N = 88). Top Scoring Pair and k-Top Scoring Pair models were evaluated against standard machine-learning classifiers and fixed EEG-ratio controls across amplitude-decile and PSD/bandpower representations using repeated subject-wise cross-validation with training-fold-only model selection. Spectral representations most consistently supported relative-ordering models. Across six prespecified disease-versus-control task–condition settings, the main k-TSP models were broadly competitive with the strongest setting-specific standard machine-learning comparators (macro-F1 differences −0.030 to +0.035). Age and sex were examined in complementary sensitivity analyses; results were moderately sensitive to demographic adjustment, whereas repeated matching produced no consistent directional loss. In regional spectral models, slow–fast restriction gave positive macro-F1 point differences in all six settings (mean +0.029), whereas removing slow–fast or theta–alpha candidates gave lower macro-F1 point estimates in all six. Participant-level permutation localised the strongest recurrence evidence at the level of broader prespecified spectral–regional patterns (motifs): these motifs exceeded corrected nulls in specific AD-related settings, although no individual exact pair did so. In a limited no-refit test on the independent ADSZ dataset (AD versus healthy controls), AD-versus-CN-trained k-TSP models reached balanced accuracy up to 0.846. These results support compact relative-ordering models as a transparent complement and favour interpretation at the broader spectral–regional-pattern level rather than from individual recurrent pairs.