⚡ electrical engineering

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

The paper introduces LEAD, the first large-scale foundation model for Alzheimer's disease detection that leverages a newly curated corpus of 2,238 subjects and a specialized gated temporal-spatial Transformer with subject-regularized training to achieve state-of-the-art generalizability and performance across heterogeneous EEG data.

Original authors: Yihe Wang, Nan Huang, Nadia Mammone, Marco Cecchi, Xiang Zhang

Published 2026-02-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yihe Wang, Nan Huang, Nadia Mammone, Marco Cecchi, Xiang Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Finding a Needle in a Haystack

Imagine trying to find a specific type of needle (Alzheimer's disease) in a massive, messy haystack (human brain activity). For a long time, doctors have had to look at tiny, individual strands of hay one by one using a magnifying glass (manual feature extraction) or a basic flashlight (standard deep learning). This is slow, often misses the needle, and doesn't work well if the haystack looks slightly different (different patients or recording machines).

The authors of this paper built a super-smart, all-seeing robot called LEAD. Instead of looking at one strand at a time, LEAD has studied millions of strands from thousands of different haystacks to learn what a "needle" actually looks like, no matter how the hay is arranged.

The Three Big Problems They Solved

The paper identifies three main reasons why previous attempts to detect Alzheimer's using EEG (brain wave sensors) have struggled:

Not Enough Data: Previous studies only looked at small groups of people (like 20 to 100). It's like trying to learn a language by only reading a few pages of a book.
- The Fix: The team gathered the world's largest collection of brain wave data for this specific problem, involving 2,238 people and hundreds of hours of recordings. They built a massive library to train their robot.
The "One-Size-Fits-All" Problem: Every person's brain is wired slightly differently, and every hospital uses different machines with different settings. A model trained on one person often fails on another.
- The Fix: LEAD is designed to be flexible. It can handle recordings of any length, any number of sensors (channels), and any speed (sampling rate). It's like a universal translator that understands every dialect of brain waves.
The "Group vs. Individual" Confusion: Most computers try to guess if a single second of brain wave is sick or healthy. But a person is sick or healthy as a whole, not just for one second.
- The Fix: LEAD focuses on the whole person. It looks at all the data from one person, aggregates the clues, and makes a final verdict on that individual, rather than guessing on every tiny slice of time.

How LEAD Works: The "Gated" Brain

The paper describes the robot's brain as a "Gated Temporal-Spatial Transformer." Here is what that means in plain English:

The Patchwork Quilt: The robot breaks the long stream of brain waves into small "patches" or pieces, like cutting a long quilt into squares.
The Three ID Tags: To understand each piece, LEAD adds three special tags:
1. Time Tag: Where does this piece fit in the timeline?
2. Location Tag: Where is this sensor on the head? (Using a 3D map of the brain).
3. Speed Tag: How fast was this machine recording?
The Gated Switch: This is the clever part. The robot looks at the data in two ways:
- Time-wise: How does the signal change over seconds?
- Space-wise: How do different parts of the brain talk to each other?
- The Gate: A smart "switch" (the gate) decides which view is more important for that specific moment. Sometimes the time pattern matters more; sometimes the location matters more. The gate learns to balance them automatically.

The Training Strategy: Learning from the Crowd

To make LEAD truly smart, the authors used a special training method:

Medical Contrastive Learning: Imagine showing the robot two pictures of the same person and asking, "Are these the same person?" Then showing two pictures of different people and asking, "Are these different?" The robot learns to recognize the unique "fingerprint" of a healthy brain versus a brain with Alzheimer's, even if the pictures look slightly different due to noise.
Subject-Regularized Training: The robot is forced to remember that all the data from one specific person belongs to that person. If the robot guesses "Healthy" for one second of a patient but "Sick" for the next second of the same patient, it gets a penalty. This forces it to be consistent about the whole person.
The Shuffle: To make sure the robot doesn't get lazy, the data is shuffled in a special way so that it always sees multiple samples from the same person in the same "batch" of training. This helps it learn the person's specific patterns better.

The Results: Beating the Competition

The team tested LEAD on 5 different groups of patients (datasets) that it had never seen before. They compared it against:

Old-school manual methods.
Standard deep learning models.
Other famous "Foundation Models" (AI models trained on huge amounts of data).

The Outcome:
LEAD came out on top in 18 out of 20 different tests. It was significantly better at correctly identifying patients with Alzheimer's compared to the other methods. Even though it was trained on less data than some of the other giant models, its smart design and specific training for Alzheimer's made it the winner.

What It Can't Do Yet (Limitations)

The paper is honest about where the robot still struggles:

The Gray Area: While LEAD is excellent at telling the difference between "Healthy" and "Alzheimer's" (or "Alzheimer's" vs. "Frontotemporal Dementia"), it is less accurate at distinguishing the early stages (Mild Cognitive Impairment or MCI) from full Alzheimer's or from being perfectly healthy.
Why? The paper explains that the brain changes gradually. A person might be in a "gray zone" where they have some Alzheimer's features but haven't fully developed the disease yet. It's like trying to tell if a seed is a sprout or a full tree when it's just starting to break the soil. The paper suggests that future work needs to track patients over many years (longitudinal data) to solve this.

Summary

In short, the authors built LEAD, a specialized AI robot that uses the largest collection of brain wave data ever assembled to detect Alzheimer's. It learns to look at the whole person, adapts to different recording machines, and uses a smart "gate" to decide what information matters most. It currently works better than any other existing method for spotting the disease, though it still finds the very early stages a bit tricky.

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

Problem Statement

Electroencephalography (EEG) offers a non-invasive, accessible, and cost-effective approach for detecting Alzheimer's Disease (AD). However, existing detection methods face three critical limitations that hinder their real-world clinical deployment:

Data Scarcity: Large-scale, high-quality EEG datasets for AD are rare. Most prior studies rely on small cohorts (20–100 subjects) and single datasets, limiting the ability to learn robust representations.
Poor Generalizability: Models often fail to generalize across unseen subjects (subject-independent evaluation). High performance in controlled, subject-dependent settings frequently collapses when faced with inter-subject variability and confounding factors like age and gender.
Heterogeneity: Existing methods struggle to adapt to highly heterogeneous data, including variations in recording environments, acquisition devices, channel layouts, sampling rates, and recording durations.

Methodology

The authors propose LEAD (Large-scale EEG model for AD detection), the first foundation model designed specifically for subject-level AD detection. The methodology is built upon a curated corpus and a novel architectural and training framework.

1. Data Curation

The authors constructed the world's largest EEG-AD corpus to date, comprising 2,238 subjects and 427.81 hours of recordings from 9 heterogeneous datasets (8 public, 1 private). For pre-training, they expanded the resource to include 9 non-AD neurological disorder datasets (e.g., Parkinson's, Depression, ADHD), totaling 4,646 subjects and 1,185.84 hours of data.

2. Architecture: Gated Temporal-Spatial Transformer

To handle arbitrary EEG configurations, LEAD employs a specialized Transformer backbone:

Patch Embedding: Slices EEG signals into univariate patches to handle variable recording lengths.
3D Channel Embedding: Utilizes sinusoidal positional encoding based on the 3D coordinates of electrodes (following the 10-20 system) to capture spatial relationships, allowing adaptation to arbitrary channel montages.
Sampling Rate Embedding: A learnable embedding explicitly informs the model of the input sampling rate, enabling adaptation to diverse acquisition frequencies.
Parallel Temporal-Spatial Attention: Applies self-attention independently along temporal and channel dimensions.
Gated Fusion: A learnable gating mechanism adaptively fuses temporal and spatial features, controlling their relative importance.

3. Training Strategy

Subject-Regularized Training: To address the gap between sample-level and subject-level performance, the authors introduce a subject-level cross-entropy loss ( $L_{sub}^{ce}$ ). This loss propagates errors at the subject level back to the model, enforcing consistency across all samples from the same individual.
Index Group-Shuffling: A novel algorithm ensures that training batches contain sufficient samples from the same subject, facilitating effective subject-level learning even with large datasets.
Multi-Sampling Segmentation: Data is downsampled to multiple rates (200Hz, 100Hz, 50Hz) and segmented into varying window lengths, increasing sample diversity and enabling multi-scale representation learning.
Domain-Inspired Pre-training: The model undergoes self-supervised pre-training using medical contrastive learning. This combines:
- Sample-level contrastive learning: Distinguishing augmented views of the same sample.
- Subject-level contrastive learning: Treating all samples from the same subject as positive pairs to learn subject-specific representations.

Key Contributions

Largest EEG-AD Dataset: The curation of a dataset with 2,238 subjects, significantly surpassing the scale of previous works (typically <100 subjects).
First EEG Foundation Model for AD: The development of LEAD, a foundation model capable of end-to-end subject-level detection that adapts to arbitrary recording lengths, channel topologies, and sampling rates.
Superior Performance: The demonstration of state-of-the-art (SOTA) subject-level results on 5 downstream AD datasets under challenging subject-independent cross-validation, outperforming both supervised deep learning methods and existing EEG foundation models.

Experimental Results

LEAD was evaluated on 5 downstream AD datasets (ADFSU, ADFTD, ADSZ, APAVA, CNBPM) against 16 baselines, including manual feature extraction, supervised deep learning models, and 4 SOTA EEG foundation models (BIOT, LaBraM, CBraMod, CSBrain).

Overall Ranking: LEAD achieved the best average ranking across all 20 evaluations (5 datasets × 2 levels: sample and subject).
Subject-Level Detection: LEAD consistently outperformed competitors in subject-level F1 scores and AUROC. For instance, on the ADFTD dataset (HC vs. AD vs. FTD), LEAD achieved a subject-level F1 of 93.95%, surpassing the next best model (LaBraM at 91.14%).
Efficiency: Despite being pre-trained on ~1,185 hours of data (significantly less than LaBraM's ~2,000 hours or CBraMod/CSBrain's ~9,000 hours), LEAD demonstrated superior generalization.
Ablation Studies: Removing key components (pre-training, sampling rate embedding, subject-level loss) resulted in performance degradation of >1% in F1 scores, confirming the necessity of the proposed design choices.

Significance and Claims

The paper claims that LEAD demonstrates the effectiveness and practical potential of large-scale pre-training for EEG-based AD detection. Key takeaways include:

Generalization: The proposed subject-regularized training strategy and domain-inspired pre-training significantly improve generalization to unseen subjects, a critical bottleneck for clinical deployment.
Adaptability: The model's ability to handle heterogeneous data (different channels, sampling rates, and lengths) makes it suitable for real-world scenarios where data acquisition standards vary.
Resource Efficiency: The results suggest that carefully curated, domain-specific pre-training on moderate-scale data can outperform models trained on much larger, less targeted corpora.

The authors position LEAD as a tool to advance research in EEG-based AD detection and inspire future studies on other brain disorders, particularly by providing pre-trained checkpoints to researchers who lack the resources to train robust models from scratch.

Limitations

The paper modestly acknowledges limitations regarding dementia-stage classification. While the model performs strongly in distinguishing AD from Healthy Controls (HC) or other dementias (FTD), its ability to separate AD, Mild Cognitive Impairment (MCI), and HC remains limited (subject-level F1 ~66%). The authors attribute this to the overlapping distribution of disease stages and the ambiguity of MCI progression (some MCI patients revert to normal, while others progress to AD), suggesting a need for longitudinal EEG data to track disease progression effectively.