2026 American Academy of Neurology Abstract Website

Objective:

In this work, we present a comprehensive research program leveraging EEG foundation models to unlock the development of new EEG-based biomarkers for neurological pathologies.

Background:

Applications derived from artificial intelligence (AI) models in EEG are limited to date due to the complexity of the EEG signal and the small volumes of accessible annotated data, incompatible with the needs of deep learning.

Design/Methods:

In order to unlock this AI limitation, we developed self-supervised EEG foundation models for clinical applications. The models combine spectral tokenization with masked token prediction to learn generalizable spatiotemporal patterns from large volumes of unlabeled EEG data. The model's ability to improve the performance of AI algorithms on various EEG use cases has been tested on numerous downstream tasks.

Results:

Our model was evaluated on three diagnostic tasks:

- Seizure detection on public Temple University EEG Seizure database (TUSZ) (n=7,152 windows)
- Normal/abnormal classification on both proprietary data (n=15,285), and public Temple University Abnormal (TUAB) database (n=23,040),
- Multiclass pathology differentiation on proprietary data (n=15,285) with 4 labels ( Normal/Lesion/Status Epilepticus/Encephalopathy).

On seizure detection task, the model achieved an AUROC at 0.926 ± 0.002 with a sensitivity at 0.909 ± 0.035. For normal/abnormal classification, the AUROC reached 0.970 ± 0.001 on proprietary database and 0.910 ± 0.002 for the public one. Finally, the multiclass pathology differentiation obtained a Weighted F1 at 0.730 ± 0.001. On these tasks our model achieved or surpassed state-of-the-art results.

Importantly, in low-data regimes (<10% of available data), the model improved AUPRC by 2–17%, demonstrating strong data efficiency.

Conclusions:

Ongoing work focuses on leveraging EEG foundation models as a universal backbone for diverse EEG diagnostic applications, including abnormality subtype classification, seizure event identification, etc.... Our findings demonstrate that foundation models can catalyze the development of versatile, resource-efficient, and clinically relevant EEG analysis systems, paving the way toward scalable AI-assisted neurology.