Morlet wavelet–based olfactory-evoked EEG features for random forest classification of normal, aMCI, and Alzheimer’s disease

Abstract

Olfactory impairment and abnormal frontal EEG oscillations are recognized as early markers of Alzheimer’s disease (AD). Using a publicly available olfactory EEG dataset of 35 subjects spanning normal cognition, amnestic mild cognitive impairment (aMCI), and AD, each with MMSE scores and demographics, stimulus-locked epochs from four electrodes (Fp1, Fz, Cz, Pz) were processed with wavelet-based time–frequency analysis. Band-limited power ratios (delta, theta, alpha, beta) were computed as log-transformed post-odor/baseline values and aggregated to subject-level features. Statistical analyses revealed graded attenuation of odor-evoked frontal (Fp1) band-power ratios across groups, with significant differences in several band–odor combinations. PCA of Fp1 features showed partial separation of diagnostic categories, while multi-channel features offered weaker discrimination. Random forest classifiers trained on Fp1-only features achieved 66.7% test accuracy, outperforming the four-channel model (55.6%), with moderate sensitivity, specificity, and precision. These findings highlight that compact frontal wavelet-derived band-power ratios during olfactory stimulation carry diagnostically relevant information for distinguishing Normal, aMCI, and AD. The transparent pipeline, combining time–frequency processing, subject-level aggregation, and multiclass classification, offers a scalable framework that can be extended to larger cohorts or integrated with multimodal biomarkers.