EEG representation model that pretrains a state-space diffusion backbone on single-channel signals, then reads its frozen latents for classification.
No providers recorded yet. Browse all providers
Most EEG foundation models borrow their pretraining recipe from language models: cut the signal into tokens, hide some, and train a transformer to fill them in. That recipe tends to reward scale, and the gains per added parameter on clinical EEG tasks have been modest. EEGDM takes the other route that self-supervised learning offers. It trains a denoising diffusion model to recover EEG from noise-corrupted copies, then treats the network's internal activations, not its output, as the representation. A model that can denoise a spike-and-wave discharge has to encode what one looks like, and those intermediate activations turn out to separate clinical event classes well.
The framework has two parts. The pretrained backbone, SSMDP (structured state-space model for diffusion pretraining), is trained once and then frozen. A lightweight latent fusion transformer (LFT) is trained on top of it for each downstream task, so a single backbone checkpoint can serve several classifiers at the same time. EEGDM was developed by Jia Hong Puah, Sim Kuan Goh and colleagues at Xiamen University Malaysia, with clinical collaborators at Universiti Malaya and co-authors at Nanyang Technological University and HKUST (Guangzhou). It was posted to arXiv in August 2025, and the repository later renamed it "EEGDM: Label-Efficient EEG Representation Learning via Generative Diffusion Model".
An unrelated 2025 model from Tsinghua University is also called EEGDM; that one uses a latent Diffusion Transformer with a ViT encoder and refers to this model as "Puah-EEGDM-SSM".
SSMDP follows the DiffWave design with 20 residual blocks, each carrying 128 gate and 128 filter channels. Every channel owns a trainable bidirectional S4D layer with a 128-dimensional hidden state. Input amplitudes pass through µ-law companding (µ = 255) to tame EEG's extreme values, and the diffusion process uses 50 timesteps with a cosine noise schedule and velocity prediction. The backbone was pretrained for 100 epochs on single-channel signals from the TUEV training set, band-pass filtered at 0.1–75 Hz and resampled to 200 Hz. For downstream use, signals pass through the backbone without added noise. Gate activations are standard-deviation pooled over one-second windows and fed to the 6.9M-parameter LFT: 20 decoder blocks for fusion and 8 encoder blocks for classification, with 8 heads and a 128-dimensional embedding. The full model, backbone plus head, is 12.8M parameters, and training ran on eight RTX 4090 GPUs.
On six-class TUEV event classification, EEGDM reports Cohen's kappa 74.23, balanced accuracy 75.57 and weighted F1 86.88. The strongest baseline, Gram-L at 251M parameters, scores 71.30, 74.87 and 88.24, so it keeps the lead on weighted F1. On CHB-MIT pediatric seizure detection, a dataset the backbone never saw, the paper reports balanced accuracy 83.79, AUPRC 48.07 and AUROC 88.78, against 73.98, 36.89 and 88.92 for CBraMod. The repository later added a transfer to IIIC ictal-interictal-injury continuum classification, with kappa 52.18 against author-reproduced baselines. That result is not yet in the arXiv paper.
EEGDM is aimed at clinical EEG review, where expert labels are scarce and expensive: classifying spike-and-wave complexes, periodic discharges, eye movements and artifacts in routine recordings, and detecting seizures in long-term monitoring. Released checkpoints cover TUEV event classification, CHB-MIT seizure detection and IIIC classification. The generative backbone can also sample synthetic single-channel EEG.
EEGDM joins the generative alternatives to masked modeling for EEG, alongside Tsinghua's unrelated model of the same name, and it makes the case at the small end of the scale. A 12.8M model pretrained on one clinical corpus matches or beats transformer foundation models such as LaBraM, EEGPT and NeuroLM on TUEV. The largest, NeuroLM-XL, has 1.7B parameters, and all three were pretrained on far more varied data. The limits are just as real. The pretraining corpus is TUEV alone, so claims about breadth rest on two transfer datasets. The work is an unreviewed preprint, and the authors describe the codebase as still under active cleanup. The code is public on GitHub but ships without a license file, so only the pretrained backbone and task checkpoints, released on Hugging Face under MIT, carry explicit terms of use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.