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

Abstract

Introduction

Localizing the epileptogenic zone (EZ) using Stereo EEG (SEEG) is often challenging through manual analysis. Even methods based on signal analysis have limitations in identifying the EZ, particularly in patients with neocortical epilepsy.

Methods

We developed machine learning (ML) methods that utilize HFO from SEEG recordings to train models to localize the EZ. We used data from 52 epilepsy patients (37 seizure free and 15 non-seizure free) who had epilepsy surgeries at our centre and were followed up for an average of 27.4 months. A total of 27 features encompassing statistical, linear, and nonlinear parameters were computed for HFOs from EZ and non-EZ brain areas. Performances of different classification algorithms were compared.

Results

In cases of mesial temporal lobe epilepsy, we achieved a cross-validation accuracy of 85.4% with the Extra-Trees classifier, 85.3% with the Random-Forest, and 82.1% with the Voting-classifier, using training data from ripples and fast ripples. For neocortical epilepsy patients, the extra trees classifier yielded an accuracy of 84.2%, while the random forest and voting classifiers attained accuracies of 84% and 80%, respectively.

Conclusion

In our approach, we employed a more realistic strategy by training the ML models at the SEEG contact level. This ensured that HFO data from a specific contact used for training the model was excluded from testing, thereby minimizing bias. This approach provides a more practical and applicable method for real-world use. Our findings indicate that the ML model-based localization of the EZ could function as an independent approach, potentially reducing the bias associated with visual analysis of SEEG.