1004 Machine Learning Epileptogenic Zone Localization Based on Spectral Power Analysis

INTRODUCTION: Surgical removal of the epileptogenic zone (EZ) after iEEG evaluation is the most successful approach for achieving seizure control in drug-resistant patients. Although automated algorithms for EZ localization hold potential, issues like technical complexity and reproducibility hurdles limit their clinical adoption. METHODS: We developed a module called FREEZ (Frequency Range Explorer to assist Epileptogenic Zone localization) within the open-source platform RAVE, enabling visualization and quantification of spectral power changes. Our study focused on assessing the predictive capability of spatiotemporal heatmaps of electrode mean power, using multi-taper Fourier transform, across six frequency ranges (delta, theta, alpha, beta, gamma, high gamma). Six random forest models were trained on seizure-free patients across these frequency ranges and combined into an ensemble, tested on both seizure-free and other patients. RESULTS: Our study included 21 ECoG patients, 14 seizure free and 7 non-seizure free, from the public fragility database. The threshold with the largest number of true positives and a positive predictive value = 95% for the random forest stacked meta-learner model output probabilities was found to be 90.67%. Prediction of electrodes within the EZ for seizure free patients by the random forest stacked meta-learner model resulted in 22 true positive predictions within the EZ, with 1 false positive predicted. True positives were found in 8 out of the 14 seizure free patients. Subsequently this trained ensemble model was tested on the non-seizure free patients, yielding 2 and 3 identified EZ electrodes that were respectively removed and non-removed. CONCLUSIONS: Our open-source FREEZ module provides efficient spectral power analysis of multi-patient IEEG data and allow the training of an efficient ensemble machine learning model for EZ electrode classification.