BEYOND-AI is a research project aimed at improving the diagnosis and surgical treatment of drug-resistant epilepsy using artificial intelligence applied to brain recordings.
About one third of people with epilepsy do not respond to medication and may need brain surgery to remove the epileptogenic zone (EZ), the brain area that generates seizures. The success of surgery depends on accurately identifying this zone. Currently, medical doctors use non-invasive methods such as EEG and magnetoencephalography (MEG), and then often implant electrodes inside the brain (stereo-EEG, or SEEG) to obtain more precise information. SEEG is highly accurate but invasive, risky, and costly.
The goal of BEYOND-AI is to make non-invasive MEG as informative as invasive SEEG, by using AI to transfer knowledge from SEEG to MEG.
The project uses exceptionally large and well-labeled datasets from Timone Hospital in Marseille: ~250 patients with SEEG recordings ~300 patients with MEG recordings ~50 patients with simultaneous SEEG and MEG, providing a unique ground truth.
The core idea is to train deep learning models on SEEG, which has a very high signal-to-noise ratio and precise expert labels, to learn subtle “hidden” patterns of epileptic activity during the interictal state (between seizures). These learned representations are then transferred to MEG using transfer learning and domain adaptation, allowing MEG to detect epileptogenic networks that are not visible to the human eye.
The project will use advanced AI methods including self-supervised and supervised deep learning, graph neural networks and multimodal data fusion. The system will produce probabilistic brain maps of the epileptogenic zone and also predict surgical outcome (e.g., likelihood of seizure freedom after surgery).
Clinically, this could reduce the need for invasive SEEG, improve surgical planning, lower risks and costs for patients. Scientifically and industrially, the project will generate new biomarkers, new AI methods for brain data, and support the development of certified clinical software.
About one third of people with epilepsy do not respond to medication and may need brain surgery to remove the epileptogenic zone (EZ), the brain area that generates seizures. The success of surgery depends on accurately identifying this zone. Currently, medical doctors use non-invasive methods such as EEG and magnetoencephalography (MEG), and then often implant electrodes inside the brain (stereo-EEG, or SEEG) to obtain more precise information. SEEG is highly accurate but invasive, risky, and costly.
The goal of BEYOND-AI is to make non-invasive MEG as informative as invasive SEEG, by using AI to transfer knowledge from SEEG to MEG.
The project uses exceptionally large and well-labeled datasets from Timone Hospital in Marseille: ~250 patients with SEEG recordings ~300 patients with MEG recordings ~50 patients with simultaneous SEEG and MEG, providing a unique ground truth.
The core idea is to train deep learning models on SEEG, which has a very high signal-to-noise ratio and precise expert labels, to learn subtle “hidden” patterns of epileptic activity during the interictal state (between seizures). These learned representations are then transferred to MEG using transfer learning and domain adaptation, allowing MEG to detect epileptogenic networks that are not visible to the human eye.
The project will use advanced AI methods including self-supervised and supervised deep learning, graph neural networks and multimodal data fusion. The system will produce probabilistic brain maps of the epileptogenic zone and also predict surgical outcome (e.g., likelihood of seizure freedom after surgery).
Clinically, this could reduce the need for invasive SEEG, improve surgical planning, lower risks and costs for patients. Scientifically and industrially, the project will generate new biomarkers, new AI methods for brain data, and support the development of certified clinical software.
Supervisor
Christian Bénar, Institut de Neurosciences des Systèmes, Aix-Marseille Université
Co-Supervisor
Mustapha Ouladsine, Laboratoire d'Informatique et Systèmes, Aix-Marseille Université
Intersectoral partner
Stefan Rampp, Uniklinikum Erlangen, Epilepsy centre, Germany
International partner
Karim Jerbi, University of Montreal, Psychology Department, Canada