An interpretable deep learning diagnostic framework for early Alzheimer’s disease based on EEG microstate spectra and multi-branch CNN

The early diagnosis of Alzheimer’s disease (AD) and its preclinical stage, mild cognitive impairment (MCI), remains challenging because of the limited diagnostic efficacy of traditional EEG microstate temporal features and the generally poor interpretability of deep learning models. This study aims to explore the frequency-domain information of EEG microstates and develop an interpretable deep learning framework to enhance early diagnostic performance and elucidate the underlying neural mechanisms. A total of 30 healthy controls (HCs), 29 individuals with MCI, and 29 individuals with AD were included in this study. We propose an interpretable deep learning framework based on frequency-domain EEG microstate information. The framework first extracts the marginal spectra of whole-duration microstate sequences across five frequency bands (the delta, theta, alpha, beta, and gamma bands) using multivariate empirical mode decomposition and the Hilbert–Huang transform, thereby constructing novel microstate-spectral features. A five-branch convolutional neural network was then designed to process spectral inputs from the frontal, parietal, central, temporal, and occipital lobes separately. Model interpretability was analyzed using Shapley additive explanations (SHAP) and gradient-weighted class activation mapping (Grad-CAM). The results revealed that the delta band power was the highest across all groups, whereas the gamma band power was the lowest. Marginal spectral values in multiple microstates and frequency bands were significantly greater in the AD group than in the MCI and HC groups (p < 0.05), with most comparisons exhibiting high statistical power (power > 0.8). The proposed model achieved an accuracy and F1- score of 92.3%±1.3% under tenfold cross-validation, significantly outperforming the conventional sliding-window power spectrum method (85.0%±2.1%), for which the accuracy was (85.0%±2.1%). It also demonstrated strong generalization ability on an independent public dataset, with an accuracy of 83.1%±3.0%. Interpretability analysis based on SHAP and Grad‑CAM further demonstrated that the model’s decisions relied predominantly on beta-band features originating from the occipital lobe. This study confirms the superior value of EEG microstate frequency-domain features in the early diagnosis of AD. By employing interpretable artificial intelligence techniques, we revealed the neural basis of the model’s decision-making, highlighting the critical role of occipital beta-band rhythmic activity. These findings provide a new methodology for developing reliable auxiliary diagnostic tools and deepen our understanding of the neural mechanisms involved in AD.