TF-JointMAE: a self-supervised multi-representation EEG learning framework for Alzheimer's disease spectrum...

Accurate identification of Alzheimer's disease (AD) and its early stages, namely subjective cognitive decline (SCD) and mild cognitive impairment (MCI), is crucial for timely intervention. Electroencephalography (EEG) is widely used for AD assessment due to its non-invasiveness and high temporal resolution; however, its non-stationarity, noise interference, and individual variability make classification more difficult. To address this, this paper proposes a self-supervised learning framework, TF-JointMAE (Temporal-Frequency Joint Masked Autoencoder), that jointly models temporal and time-frequency representations of EEG and incorporates age information as a conditional physiological prior within a unified embedding space. By performing self-supervised masked autoencoder pre-training on multiple public EEG datasets, the model learns consistent and robust EEG representations, thereby improving AD-spectrum classification under limited-label conditions. On the publicly available CAUEEG dataset (Normal, SCD, MCI, Dementia) and the olfactory EEG dataset (Normal, MCI, Dementia), TF-JointMAE achieved test accuracies of 77.97% and 96.43%, respectively, and demonstrated higher discriminative stability in the MCI category. Further occlusion-sensitivity analysis revealed that the model showed varying sensitivity to EEG channels and time-frequency regions across cognitive states. These results demonstrate that TF-JointMAE effectively improves the robustness of EEG representations, providing potential auxiliary support for AD-spectrum classification and clinical decision-making. All source code is publicly available to support reproducibility (https://github.com/redpig-zhu/TF-JointMAE).

The online version contains supplementary material available at 10.1007/s11571-026-10501-8.