Authors :
Presenting Author: Aurore Semeux-Bernier, PhD Student – Institut de Neurosciences des Systèmes, AMU
Francesca Bonini, MCU-PH – AP-HM, Timone Hospital; Maria Fratello, Research engineer – INSERM, Institut de Neurosciences des Systèmes, AMU; Elodie Garnier, Research engineer – INSERM, Institut de Neurosciences des Systèmes, AMU; Samuel Medina Villalon, Research engineer – INSERM, Institut de Neurosciences des Systèmes, AMU; Jean-Michel Badier, IR AMU – INSERM, Institut de Neurosciences des Systèmes, AMU; Frédéric Richard, PR – CNRS, Institut de Mathématiques de Marseille, AMU; Christian-George Bénar, DR – INSERM, Institut de Neurosciences des Systèmes, AMU
Rationale:
In epilepsy, to cure pharmaco-resistant patients, one solution consists in resecting the brain areas responsible for seizures. To determine this area, clinicians use semiology, neuroimaging analysis complemented with non-invasive recording like electroencephalography (EEG), magnetoencephalography (MEG) (Laohathai, et al. 2021) (Phase 1) and subsequently invasive recording with stereo-encephalography (SEEG) (Bartolomei, et al. 2018) (Phase 2). The position of SEEG electrodes can be guided by the non-invasive signal analysis. However, the surgery is failing in 30-40% of the cases (Owen, et al. 2023). Functional constraints exclude, it could mean that the current analysis is not optimal. It is particularly difficult to extract relevant information from the interictal activity of EEG/MEG. Our study aims to enhance the analysis of MEG by finding biomarkers relevant in the assessment of the epileptogenic zone (EZ)
Methods:
We computed Independent Component Analysis (ICA) (Iriarte, et al. 2006) on MEG data for 30 patients. Assuming that one IC corresponds to one source in the brain, we aimed to classify each component between epileptic and non-epileptic using logistic regression. We trained and tested our model relying on the labels (epileptic or not) of ICs assigned by an expert.
Results:
We found that even if there is a significant difference on the distribution of features between epileptic and non-epileptic ICs, this does not separate efficiently the two classes (mean F1-score = 0.38). Still, there is a significant difference in classification from chance level (mean F1-score = 0.12, p ≤ 0.05). The presence of spikes is a good marker for a clinician to visually recognize an epileptic IC, but it is more difficult in an automatic mode (mean F1-score = 0.23). Yet, the combination of features is improving the classification (p ≤ 0.05).
Conclusions:
These results are encouraging as they show the potential of using and combining features to classify ICs and to find epileptic sources. However, we would have needed more subjects and maybe more complex features. Moreover, only logistic regression was tested but many other models of classification exist that ought to be tried, such as non-linear models for example.
Funding:
ASB is supported by a Laennec PhD grant, the project has received funding from the Excellence Initiative of Aix-Marseille Université - A*Midex, a French “Investissements d’Avenir programme” AMX-21-IET-017 and France Life Imaging network (grant ANR-11-INBS-0006).