Abstracts

Rationale: Some patients with refractory focal epilepsy are chronically implanted with a responsive neurostimulation device (RNS System, NeuroPace, Inc., Mountain View, CA), an FDA-approved closed-loop system with delivers electrical stimulation upon detection of an abnormal pattern in the electrocorticography (ECoG) . Physicians program RNS detection and stimulation parameters to potentially reduce seizure duration and frequency. However, current clinical practice is limited by a lack of evidence and biomarkers to guide stimulation settings, leading to inefficient progress toward optimizing therapy or potentially worsening seizure outcomes. Our goal is to build a reliable classifier for individual patients to predict seizure frequency in a given month, by applying machine learning approaches to the RNS Background ECoG. Methods: We identified epilepsy patients implanted with the RNS System at a single center with: 1) More than 6 months from implantation, (2) good RNS upload compliance, as defined by upload on more than 2/3 of days, (3) good clinical correlation between long episodes and clinical seizures, and (4) scheduled background RNS EcoGs at least twice per day. We analyzed their retrospective ECoG data from extended periods (> 200 days) with unchanged RNS detection parameters and antiepileptic drug regimens. This extended window was then further divided into 1-month epochs. Mean daily long-episode (LE) counts, a surrogate for seizure frequency, was determined for each epoch. Each epoch was classified as “Good” if mean LE count was below the median, and “Bad” if otherwise. All ECoG segments were pooled within each condition to form a binary dataset, e.g.the “Good” class contained all scheduled ECoGs from all “Good” epochs. Simple band power features were calculated for classic frequency bands within the hardware filter bandpass (4-90 Hz) for each recorded ECoG segment. Basic machine learning algorithms, including logistic regression, random forest, gradient boosting, SVM etc. were systematically trained on 80% of the entire dataset based on the spectral features. Ten-fold cross-validation was applied to grid-search for optimal hyperparameters. The resulting classifiers were then tested on the remaining 20% of data. We then analyzed the effect of arousal state on classification accuracy by separating awake and sleep ECoGs. An additional three classifiers were trained with 1) only sleep ECoGs, 2) only awake ECoGs, and 3) with all ECoGs with an additional binary feature of sleep versus awake. Results: To date, we have analyzed four patients with the RNS System. Gradient boosting and logistic regression approaches appear to yield the best performance, achieving an area under the ROC curve (AUC) from 0.7 (fair) to 0.9 (excellent). Feature ranking analysis showed that higher frequency bands (beta, low and high gamma) tended to be the most important features in classification, with decreased power associated with good epochs. Two patients with suitable data for wake/sleep ECoG analysis revealed that classifiers trained with only wake or only sleep ECoG (depending on what time of day the patient had seizures), achieved better or comparable classification performance compared to utilizing all samples. Conclusions: Machine learning algorithms can be applied to retrospectively analyze background RNS ECoG segments of individual patients, to yield models which classify seizure frequency with fair to excellent performance. Gradient boosting and logistic regression algorithms appear to outperform other approaches. Higher frequency activity (beta, low gamma, high gamma) may be an important predictive feature. Time of day when patients have seizures (daytime wake/nighttime sleep) may be valuable in classification performance. Funding: NIH NINDS K23NS104252 (LIU), NYU Moore Sloane Data Science Grant