Abstracts

Rationale: The apparently unpredictable timing of seizures is a major factor limiting quality of life in patients with epilepsy. Previously, responsive neurostimulator (RNS) data have been used to forecast seizure probability using detections of clinician-programmed patterns. However, to our knowledge models predicting seizures have thus far not incorporated the details of individual detected interictal events. With the goal of informing such models, we sought to characterize which features of RNS data – including pattern type detected, episode duration, and therapies delivered – preceded and correlated with seizure onset.

Methods: From a database of 53 patients with RNS who have been seen at one center, we included patients who met the following strict criteria: an epoch of constant pattern detection settings greater than or equal to 6 months, no more than 50% of days with seizures, and at least an average of 20 seizures every 6 months, with data recorded from at least 90% of events. This search yielded an initial dataset of 8 patients (5 female). We then characterized individualized seizure surrogate markers (e.g. long episodes or saturations) for each patient and annotated RNS data to mark events meeting seizure criteria. Data were resampled to generate time series with median event duration in seconds, total number of long episodes, total patterns detected, total therapies delivered, total events with number of therapies exceeding a threshold (e.g. events with more than 2 therapies delivered), and presence or absence of a seizure surrogate calculated over 10-minute intervals. Each RNS feature was cross-correlated with seizures in each epoch, and the peak normalized cross-correlation was identified.

Results: We found that for 6 of 8 patients, long episodes demonstrated the greatest peak cross-correlation with seizure onset out of features examined. For the remaining 2 patients, the features with greatest peak cross-correlation with seizures were a clinician-programmed pattern and episode duration. For 4 of 8 patients, total events with number of therapies exceeding a threshold demonstrated greater peak cross-correlation with seizures as the threshold increased (e.g. greater peak cross correlation with seizures for events with 2 or more therapies delivered than for events with 1 or more).

Conclusions: These findings illustrate features of RNS data temporally correlated with seizure onset, which may help inform predictive models for seizure forecasting using RNS data.

Funding: Please list any funding that was received in support of this abstract.: James G. Hirsch Endowed Medical Student Research Fellowship (EMY) and the Swebilius Foundation (IHQ).