Improving the accuracy of epileptogenic zone localization in stereo EEG with machine learning algorithms

The precise identification of the epileptogenic zone (EZ) is paramount in the presurgical evaluation of epilepsy patients to ensure successful surgical outcomes. The analysis of Stereo EEG, an instrumental tool for EZ localization, poses considerable challenges even for experienced epileptologists. Consequently, the development of machine learning (ML)-based computational tools for enhanced EZ localization is imperative. In this investigation, we developed ML models utilizing Stereo EEG from 15 patients, who remained seizure-free (Engel 1 a-d) following EZ resection, over an average follow-up period of 44.4 months. Utilizing Delphos and MNI detectors, spikes and High Frequency Oscillations (HFOs) were identified from Stereo EEG in Resected Zone (RZ) and non-Resected Zone (non-RZ). Linear and non-linear features were estimated from each modality using MATLAB. A total of 27,744 spikes, 7,790 ripples, and 7,632 fast ripples, along with their combinations, were employed to train the ML models. The Gradient Boosting classifier demonstrated the highest prediction accuracy of 98.5% for EZ localization in Mesial Temporal Lobe Epilepsy (MTLE) when trained with features derived from the spike-ripple combination. In the case of Neocortical Epilepsy (NE), the Extra Trees classifier achieved an accuracy of 87.6% when utilizing features from fast ripples. The Random Forest, Extra Trees, and Gradient Boosting algorithms were the most effective for predicting the RZ. Linear features outperformed non-linear features in predicting epileptogenic zones within the epileptic brain. Our study establishes the capability of ML methodologies in localizing epileptogenic zones with high accuracy. Future studies that focus on increasing the training sample size and incorporating more advanced machine learning (ML) algorithms have the potential to significantly improve the accuracy of these models in pinpointing epileptogenic networks. Additionally, implementing this ML approach across multiple research centers would contribute to the broader validation and generalizability of this technique.