Thirty New Personalised ‘Virtual Brain Twins’ Offer Ground Truth for Epilepsy Research

Searching for the source of seizures

About one-third of epileptic patients live with drug-resistant epilepsy. For these patients, brain surgery is often the best chance of a cure but it fails in roughly 50% of cases, mainly because precisely identifying the specific brain region responsible for triggering seizures, the “Epileptogenic Zone” (EZ), is challenging. 

While clinicians typically use Stereo-electroencephalography (SEEG), by inserting depth electrodes into the brain, to record electrical activity and define the EZ, finding the EZ remains difficult because the exact seizure origin is never fully known for empirical data. 

To overcome this challenge, neuroscientists at the Institut de Neurosciences des Systèmes (Marseille, France) have created and released a new cohort of 30 "Virtual Epilepsy Patients", using a brain modelling approach - based on individual patients’ data - called Virtual Brain Twin.

Methods: Building the Virtual Cohort

The researchers used the Virtual Epileptic Patient (VEP) pipeline to create personalised whole-brain models for 30 real patients with drug-resistant epilepsy, resulting in a comprehensive dataset. This dataset includes structural brain data, model parameters, and simulated synthetic SEEG time series for three critical brain states: spontaneous seizures, stimulation-induced seizures, and interictal (between-seizure) activity. 

1. Digital replication: They built a digital replica of each patient's brain using MRI scans alongside the VEP atlas to map the brain’s anatomy and structural connectivity of different regions. 

2. Seizure modelling: They applied the Epileptor model, which uses advanced mathematics to simulate the brain activity of these regions, and describes spatio-temporal seizure dynamics  (how seizures move over space and time).

3. Establishing ground truth: The ground truth is defined by the model parameters used to generate the simulated brain activity at the SEEG level. This simulated activity is performed at the whole brain level: each region generates local mean brain dynamics of the population of neurons surrounding it. The patient’s estimated EZ was used to set the model’s excitability parameter for each brain region, providing a reliable reference for the simulations.

4. Generating data: Finally, the whole-brain simulation was mapped onto virtual SEEG electrodes, which were reconstructed from patient CT scans, to generate the synthetic SEEG time series (electrical recordings).

By tuning this digital brain to behave like the real patient, researchers generated "synthetic" seizures that mimic empirical recordings. This provides a key advantage: because the underlying generative seizure parameters of these virtual brains are fully controlled and known, this synthetic cohort gives scientists a reliable "answer key" that real patient data cannot offer, allowing them to test their methods before applying them on real patients.

Validation and accuracy

The personalised Virtual Epileptic Cohort (VEC) demonstrated high accuracy when compared against intracranial SEEG recordings and outperformed a randomised cohort. This confirmed that patient-specific EZ identification and chosen model parameters are important in capturing epileptic activity.

Furthermore, when researchers simulated electrical stimulation (a procedure used in pre-surgical mapping), the models accurately predicted which regions would trigger seizures and how seizures would propagate through the brain network, matching patterns observed in real patients. These digital twins, therefore, proved to be a reliable tool for identifying the EZ and predict surgical outcomes.

Significance of key models and implications

This open-access dataset represents a key resource for the neuroscience community, impacting research across three main areas:

• Benchmarking: It provides for the first time the "ground truth" necessary to validate algorithms for source localisation and EZ estimation, a capability previously impossible with clinical recordings alone.
• Privacy and Access: Synthetic data mimics the features of real patient data without compromising patient privacy, making it easier to share across institutions and borders.
• Clinical Translation: The underlying VEP pipeline is currently used in the EPINOV clinical trial, which evaluates whether virtual brain modelling can improve surgical outcomes for drug-resistant epilepsy patients. 

By hosting this dataset on EBRAINS, the European digital infrastructure for brain research, the study showcases the platform's role in advancing transparent reproducible research by providing standardised data formats and openly accessible virtual models, and fostering collaboration between data scientists and clinicians.

Availability:

The full findings are published in the journal PLOS Computational Biology (DOI: 10.1371/journal.pcbi.1012911).  

The virtual epilepsy patient cohort is openly available and can be used to systematically evaluate methods for the estimation of EZ or source localisation. It has been uploaded to EBRAINS. The dataset card can be accessed via the following link: doi: 10.25493/ XWG2-S8X. 

The code used to generate the synthetic data and perform the analyses is also openly accessible on GitHub.

Citation:

Dollomaja B, Wang HE, Guye M, Makhalova J, Bartolomei F, Jirsa VK (2025) Virtual epilepsy patient cohort: Generation and evaluation. PLoS Comput Biol 21(4): e1012911. https://doi.org/10.1371/journal.pcbi.1012911