Abstract Accurate identification of the seizure onset zone (SOZ) using intracranial electroencephalography (iEEG) remains challenging. Although diverse methods have leveraged spectral features to classify patient outcomes, few approaches focus on identifying individual electrodes within the SOZ or integrate a broad spectrum of frequency ranges within a single model. We developed an interpretable machine learning model that integrates power across delta, theta, alpha, beta, gamma, and high-gamma frequencies over time to identify the SOZ. For 1,511 electrodes implanted across 21 patients, we computed the mean spectral power in each frequency band for the first 20 seconds after seizure onset and analyzed the differences in power between SOZ and non-SOZ electrodes. In patients who were seizure-free after surgery (n = 14), electrodes within the SOZ showed significantly higher area under the curve (AUC) for mean power over time in the first 20 seconds after seizure onset compared to electrodes outside the SOZ in the alpha (p = 0.0272), beta (p = 0.0263), gamma (p = 0.0013), and high gamma (p = 0.0086) ranges. Additionally, electrodes within the SOZ in patients that became seizure-free after surgery had significantly higher AUC compared to equivalent electrodes in patients who did not become seizure-free after surgery (n = 7) in the gamma (p = 0.0145) and high gamma (p = 0.0024) power ranges. We trained a stacked random forest ensemble model using these features over time to label electrodes within the SOZ. Leave-one-out patient cross validation of the machine learning model yielded a 96.6% positive predictive value and 99.9% specificity for identifying electrodes within the SOZ. Our dataset included a diverse array of seizure onset patterns, which were all classified accurately by the model. A second model was trained to predict post-operative seizure freedom, yielding 95.2% accuracy for predicting seizure outcome based on a planned resection. This two-model design mirrors clinical workflow, first localizing SOZ electrodes to support surgical planning, then predicting outcome based on a surgical plan. An advantage of our interpretable machine learning approach is the ability to interrogate our models to understand how predictions are made. For electrode classification, the model weighed beta (0.66 ± 0.07), high gamma (0.54 ± 0.06), and delta (0.51 ± 0.06) power bands most heavily. Viewing the model’s frequency band weights over time reveals that the model identified a pattern resembling the “fingerprint of the epileptogenic zone”, reinforcing the importance of this dominant fundamental neurophysiologic signature of seizure onset.
O’Leary et al. (Tue,) studied this question.