AI Epilepsy 2027 – Keynote: EEG-Based Foundation Models
Abstract
Invited keynote at AI Epilepsy 2027, the 5th International Conference on Artificial Intelligence in Epilepsy and Neurological Disorders, which brings clinicians, neuroscientists, engineers, and computer scientists together to move AI for epilepsy and neurology into real-world care. The talk covers EEG-based foundation models: models pre-trained on tens of thousands of hours of unlabeled EEG that transfer to clinical tasks such as seizure detection, abnormality detection, and artifact rejection with little labeled data. I will discuss how our models handle the practical realities of clinical EEG — heterogeneous electrode montages (LUNA, LuMamba), long recordings at linear cost (FEMBA, LuMamba), multimodal EEG/ECG/PPG signals (PanLUNA), and label-efficient self-supervised pre-training — and what it takes to run them on wearable devices for continuous, long-term epilepsy monitoring.
I’m honoured to have been invited to give a keynote on EEG-based foundation models on February 22, 2027, the opening day of AI Epilepsy 2027 (February 22–25) in Breckenridge, Colorado — the mountain town where the conference series began in 2023.
What I’ll cover
- Why foundation models for EEG: labeled clinical EEG is scarce and expensive, but unlabeled recordings are abundant. Self-supervised pre-training turns that archive into models that adapt to new clinical tasks with little annotation.
- Any montage, any length: clinical and wearable EEG use very different electrode layouts and recording durations. LUNA (NeurIPS 2025) reads any electrode layout into a shared latent space, FEMBA and LuMamba scale linearly with recording length, and PanLUNA extends the idea to joint EEG, ECG, and PPG.
- From the cloud to the patient: compressing these models until they run in real time on ultra-low-power wearables — the path towards continuous, long-term epilepsy monitoring outside the hospital.
Resources
Header banner: AI Epilepsy 2027 organizers.
Breckenridge in February
At 2,900 m in the Colorado Rockies, Breckenridge in late February means deep snow and groomed Nordic trails winding through the pines — a fitting setting for a meeting that started here.
, [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/).](/talk/ai-epilepsy-2027-keynote-eeg-based-foundation-models/breckenridge_nordic_trail_hud0d3dd23f15f1f634f5150d6a04a807e_212518_077c4903c4ff8ccb176fb492066007e4.jpg)
, [CC BY 2.0](https://creativecommons.org/licenses/by/2.0/).](/talk/ai-epilepsy-2027-keynote-eeg-based-foundation-models/breckenridge_nordic_center_hub525426fc415a39773ba85eedc13dcc6_100246_5a20b52aa3800eabaf4838ec20cb343f.jpg)
Thorir Mar Ingolfsson
Postdoctoral Researcher
Making AI run on microwatts: I build foundation models for EEG, ECG, and other biosignals, and compress them until they run in real time on wearable devices, not in the cloud.