Impact of Recording Conditions and Feature Fusion on Dementia Classification Using Spectral, Entropy, and Complexity EEG Features

Resting-state electroencephalography (EEG) is a promising low-cost, non-invasive modality for supporting differential classification of Alzheimer’s disease (AD), mild cognitive impairment (MCI), and healthy controls (HCs). However, it remains unclear whether the combination of different EEG feature families consistently improves classification and whether their utility varies with recording condition. We compared spectral, entropy-based, and complexity-based features, both individually and in combination, for AD-HC, AD-MCI, and MCI-HC classification under eyes-closed (EC) and eyes-open (EO) conditions using an age-matched paired subset of the CAUEEG data set. Classification was evaluated using nested leave-one-subject-out cross-validation with two-stage hierarchical LightGBM gain-based feature selection. Under the EC condition, the highest balanced accuracies (BAs) were obtained with spectral features for AD-HC (79.99%), spectral–entropy for AD-MCI (68.33%), and spectral–entropy–complexity for MCI-HC (63.15%). Under the EO condition, spectral features performed best for AD-HC (84.91%) and AD-MCI (66.90%), whereas entropy performed best for MCI-HC (70.14%). Feature fusion therefore showed task-dependent rather than consistent benefits. Although numerical EC-EO differences were observed, none remained significant after paired permutation testing. Overall, the results suggest that the relative usefulness of EEG feature families varies across diagnostic comparisons and recording conditions.