Studies suggest the existence of a pre-critical state preceding the onset of an epileptic seizure. Identifying these states from self-reported prodromal symptoms, combined with machine learning algorithms, could help anticipate seizures.

Around 65 million people worldwide, or 1% of the global population, suffer from epilepsy. It is the 3rd most common neurological pathology. Epilepsy is a chronic condition liable to generate spontaneous and repeated epileptic seizures, and it is estimated that around a third of patients are drug-resistant and will continue to have seizures despite appropriate anti-epileptic treatment. The onset of a seizure is a paroxysmal and unpredictable phenomenon - "a thunderclap in a serene sky" - which accounts for the handicap and social repercussions for patients.

The concept of a limited two-state model in epilepsy - i.e. intercritical/critical - has been challenged in recent decades. Ictogenesis could include a transitional state characterized by changes in cortical excitability that would pave the way for the onset of an epileptic seizure. This so-called pre-critical state is the scientific basis for seizure prediction models. If this state can be detected long enough before the onset of a seizure to detect a change in the brain's state, a seizure-stopping intervention (medication, biofeedback techniques, stimulation techniques, etc.), or at least safety measures, can be proposed.

While a deterministic approach has long been applied to predictive models - to predict the occurrence of the next crisis - a new strategy has more recently developed. Today's strategies are more realistic and adapted to non-linear dynamic systems. Indeed, probabilistic approaches from the meteorological sciences are increasingly being applied to crisis prediction models. The aim of crisis forecasting is to estimate the probability of a future crisis at any given time, whereas classical prediction algorithms aim to accurately predict the occurrence of a future crisis. In this way, we can identify a "pro"-critical state, i.e. a state at high risk of epileptic seizure.

Several studies have suggested the existence of a pre-critical period. However, identifying specific pre-critical biomarkers remains a major challenge. While information derived from EEG signals has long been favored, analysis of clinical symptoms has emerged more recently. Pre-critical clinical symptoms, otherwise known as "prodromes" or "prodromal symptoms", may precede the seizure by several hours. Some studies have also highlighted the value of integrating self-prediction - the patient's subjective assessment of the risk of an upcoming crisis - without anticipation models.

Previous work by the investigators has developed a classification algorithm capable of identifying a pre-critical state from the daily assessment of several prodromal symptoms. These results were obtained in a hospital setting, with good classification performance. This work was the subject of a European patent application (No. 20306548.7) on December 11, 2020 and an international patent application (No. PCT/EP2021/085146) on December 10, 2021: "A computer-implemented model for predicting occurrence of a seizure and training method thereof".

The main hypothesis of this study is that a machine learning algorithm based on the daily assessment of prodromal symptoms could identify seizure-prone states in patients with epilepsy.

Inclusion Criteria:

• Age between 18 and 65
• Focal epilepsy diagnosed for at least 18 months
• Brain imaging as part of the etiological work-up for epilepsy showing no progressive cause
• EEG compatible with the diagnosis of epilepsy within the last 10 years
• At least 2 non-contiguous days of epileptic seizures per month, according to the patient
• Ability of the patient to understand and use a mobile application on the personal smartphone
• Free, informed and signed consent
• Affiliation with a social security scheme (excluding AME)

Exclusion Criteria:

• Suspicion or diagnosis of other types of associated malaise: functional dissociative seizures, syncope or other malaise of non-neurological origin
• Assessment of seizure frequency deemed unreliable by the investigator (eg. due to cognitive impairment)
• Inability to describe seizures accurately
• Presence of more than 15 days with seizures per month
• Participation in other interventional research or exclusion period not expired
• Pregnant or breastfeeding woman
• Patient under guardianship, curatorship, deprived of liberty