Machine Learning-Based localization of the epileptogenic zone using High-Frequency oscillations from SEEG: A...

Area of Science:

• Neuroscience
• Medical Technology
• Machine Learning

Background:

• Manual analysis of Stereo EEG (SEEG) for localizing the epileptogenic zone (EZ) is challenging.
• Existing signal analysis methods have limitations, especially for neocortical epilepsy.

Purpose of the Study:

• To develop and evaluate machine learning (ML) methods using high-frequency oscillations (HFOs) from SEEG recordings for precise EZ localization.
• To compare the performance of different classification algorithms in identifying the EZ.

Main Methods:

• Developed ML models utilizing 27 features from SEEG-recorded HFOs (statistical, linear, nonlinear).
• Trained and tested models at the SEEG contact level using data from 52 epilepsy patients.
• Compared classification algorithm performance for mesial temporal lobe and neocortical epilepsy.

Main Results:

• Achieved cross-validation accuracies up to 85.4% for mesial temporal lobe epilepsy using Extra-Trees and Random-Forest classifiers.
• Attained accuracies up to 84.2% for neocortical epilepsy with the Extra-Trees classifier.
• Demonstrated high performance across multiple classifiers for both epilepsy types.

Conclusions:

• ML models trained at the SEEG contact level provide a realistic and unbiased approach to EZ localization.
• This method minimizes bias by excluding training data from testing, enhancing practical applicability.
• ML-based EZ localization can serve as an independent method, reducing reliance on subjective visual analysis of SEEG.