EEG foundation model for Alzheimer's disease detection, pretrained by contrastive learning across 13 clinical EEG datasets and 2,238 subjects.
No providers recorded yet. Browse all providers
Electroencephalography is cheap, non-invasive, and widely available, which makes it an attractive substrate for screening cognitive decline — but the datasets are small. Individual Alzheimer's disease (AD) EEG cohorts typically hold a few dozen to a few hundred subjects, so deep models trained on one cohort rarely survive contact with another site's electrodes, amplifiers, or recording protocol. LEAD attacks that problem the way foundation models attack data scarcity elsewhere: pool everything, pretrain once, and transfer.
Developed by researchers at the University of North Carolina at Charlotte with collaborators at the Mediterranea University of Reggio Calabria and the EEG diagnostics company Cognision, LEAD was released as a preprint in February 2025. The authors curated what they describe as the largest EEG corpus assembled for Alzheimer's detection — 2,238 subjects across 13 pretraining datasets, four containing AD patients and nine covering other neurological conditions or healthy controls — and pretrained a transformer on it with a medical contrastive objective. Crucially, evaluation happens on five separate AD datasets that were never seen during pretraining, making this a cross-cohort transfer test rather than a held-out split.
The design contrast with general-purpose EEG backbones such as LaBraM and CBraMod is one of specialization: those models pretrain broadly and fine-tune to many downstream tasks, while LEAD targets a single clinical decision and buys its generality from disease-adjacent pretraining data.
Input EEG is sliced into univariate patches and embedded with temporal positional, 3D channel-coordinate, and sampling-rate embeddings before entering the gated temporal-spatial transformer. Pretraining uses medical contrastive learning over the 13-dataset corpus; downstream adaptation fine-tunes on each of the five held-out AD datasets. Across the 20 evaluation settings spanning those five datasets, LEAD achieves the best average ranking, outperforming both conventional EEG classifiers and published EEG foundation models. Pretraining data is preprocessed with notch and 0.5–45 Hz band-pass filtering, average re-referencing, ICA-based artifact removal, and alignment to the standard 19-channel 10-20 montage; downstream datasets keep their native channel layout. Code and a pretrained checkpoint are distributed from the project repository under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 license, which restricts commercial deployment.
The intended user is a group with a modest Alzheimer's EEG cohort and no budget to pretrain from scratch — a memory clinic, a neurology department, or a trial site collecting resting-state recordings. Fine-tuning LEAD on a few hundred local subjects is a far shorter path to a working subject-level classifier than training a network on that cohort alone. Because the model tolerates arbitrary montages and sampling rates, it also fits retrospective analyses of archived EEG collected under protocols that have drifted over the years, and it offers a template for other EEG-based neurological screening tasks where labeled data is similarly scarce.
LEAD is an argument that the foundation-model recipe transfers to clinical biosignals where the bottleneck is patient recruitment rather than compute. Its most durable contribution may be the curated corpus and preprocessing pipeline: the repository documents 18 public and private EEG datasets with per-dataset preprocessing notebooks, which lowers the entry cost for anyone benchmarking in this space. The limitations are real and acknowledged — the model is a preprint awaiting peer review, the non-commercial license blocks product use, and cross-dataset transfer in EEG remains sensitive to site-specific acquisition artifacts that no amount of channel alignment fully removes.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.