LEAD: An EEG Foundation Model for Alzheimer's Disease Detection

TL;DR

This paper addresses the scarcity and heterogeneity of EEG data for Alzheimer's detection by building LEAD, the first large-scale EEG foundation model trained on the world’s largest EEG-AD corpus (2,238 AD-related subjects) and 13 heterogeneous datasets. LEAD uses a gated temporal-spatial Transformer with univariate patch embeddings, 3D channel embeddings, and sampling-rate embeddings, paired with subject-regularized training and domain-inspired self-supervised pre-training. Across 5 AD downstream datasets and 20 evaluations, LEAD achieves top rankings and outperforms state-of-the-art EEG foundation models, validating the effectiveness of large-scale EEG pre-training and subject-focused learning for clinical detection. The work emphasizes practical deployment potential and provides pre-trained checkpoints to support research on AD and related brain disorders.

Abstract

Electroencephalography (EEG) provides a non-invasive, highly accessible, and cost-effective approach for detecting Alzheimer's disease (AD). However, existing methods, whether based on handcrafted feature engineering or standard deep learning, face three major challenges: 1) the lack of large-scale EEG-based AD datasets for robust representation learning; 2) limited generalizability across subjects; and 3) difficulty in adapting to highly heterogeneous data. To address these challenges, we curate the world's largest EEG-AD corpus to date, comprising 2,238 subjects. Leveraging this unique resource, we propose LEAD, the first large-scale foundation model for EEG-based AD detection. Specifically, we design a gated temporal-spatial Transformer that can adapt to EEG recordings with arbitrary lengths, channel configurations, and sampling rates. In addition, we introduce a subject-regularized training strategy to enhance subject-level feature learning. We further employ medical contrastive learning for pre-training on 13 datasets, including 4 AD datasets and 9 non-AD neurological disorder datasets, and fine-tune/test the model on the other 5 AD datasets. LEAD achieves the best average ranking across all 20 evaluations on 5 downstream datasets, substantially outperforming existing approaches, including state-of-the-art (SOTA) EEG foundation models. These results strongly demonstrate the effectiveness and practical potential of the proposed method for real-world EEG-based AD detection. Source code: https://github.com/DL4mHealth/LEAD

Figures (3)

• Figure 1: LEAD overview. We pre-train LEAD on 13 datasets of neurological disorders using medical contrastive learning. The model is then fine-tuned on 5 AD datasets under the subject-independent cross-validation, with training, validation, and test splits of 80%, 10%, and 10%, respectively. We incorporate novel training strategies, including multi-sampling segmentation and index group shuffling. A gated temporal-spatial Transformer is employed to capture both temporal and spatial features. In addition, we propose a subject-level cross-entropy loss $\mathcal{L}_{ce}^{sub}$ to enhance the learning of subject-level representations. Finally, majority voting is used for subject-level detection.
• Figure 2: a) Gated Temporal Spatial Transformer. Input EEG samples are first sliced into univariate patches. These patches are mapped to patch embeddings, to which temporal positional embeddings, 3D channel embeddings, and sampling rate embeddings are added to form the final patch representations. A parallel temporal-spatial attention mechanism is then applied along both the temporal and channel dimensions. The resulting features from the two dimensions are combined using a learnable gated fusion module. b) Subject-Regularized Training. Each subject’s EEG recordings are segmented into windows with varying sampling rates, such as 200 Hz, 100 Hz, and 50 Hz, using a multi-sampling segmentation strategy. Index group shuffling ensures that each training batch contains sufficient samples from the same subject, which facilitates subject-level learning. The classifier produces sample-level predictions that are aggregated into subject-level predictions. Both two-level predictions are used to compute cross-entropy losses.
• Figure 3: a) Average performance rank of 17 methods across all 5 datasets $\times$ 2 evaluation metrics $\times$ 2 levels. For example, the value 1.5 of LEAD indicates an average rank of 1.5 over the total 20 evaluations. Lower ranks and a deeper blue indicate better performance.