Abstracts

This abstract is a recipient of the Young Investigator Award
This abstract has been invited to present during the Translational Research platform session

Rationale: Seizure clustering, i.e., when multiple seizures occur within a short duration, is estimated to occur in 13% to 76% of people with epilepsy. However, on termination of a given seizure, it is unknown whether the seizure is isolated or part of a cluster. Identifying whether a seizure is isolated is clinically important for mitigating the impact of seizure clusters. Our goal is to develop a machine learning (ML) model for predicting seizure type, i.e., isolated seizures or cluster seizure._x000D_ _x000D_ Methods: Seizure type prediction is challenging because of limited (1) availability of data suitable for prediction, and (2) understanding of the electrographic features differentiating isolated and cluster seizures. We used ML for predicting seizure type based on relative entropy (REN), a bivariate iEEG feature, extracted from physiologic frequency bands during ictal and pre-ictal periods. REN is used to quantify interactions between electrodes and capture differences in seizure types. Long-term iEEG data with sufficient number of seizures to enable the training of patient-specific prediction models was used. For each seizure, iEEG data from the ictal and 10 minutes pre-ictal period were divided into non-overlapping 2.5-s segments and REN between all pair of electrodes was extracted. Grand average of the REN for all segments and pairs of electrodes, separately for each band and period, were used as prediction features of each seizure. Seizures within 24 hours of each other were considered as a cluster._x000D_  _x000D_ Two classification tasks were considered: (1) Patient-specific random forest classifier (RFCs) predicted whether the next seizure would occur within 24 hours of the given seizure or not. For this task, seizures were divided as follows: (i) “isolated + cluster-last” - isolated seizures and the last seizure in clusters, since no seizure occurs shortly after them; and (ii) “cluster-non last” - the remaining seizures in clusters. (2) To predict seizure cluster onset, patient-specific RFCs predicted whether a seizure was isolated or the first seizure in a cluster. Only isolated seizures and the first seizure in each cluster (cluster-first) were considered for this task._x000D_ _x000D_ Results: We used data of 3710 seizures collected from iEEG recordings spanning 4951 days from 9 patients in the NeuroVista study. On average across patients, the model predicted whether another seizure would occur within 24 hours after seizure termination with 78.4% (± 17.1%) F1-score (which combines sensitivity and precision) (Table 1). The performance was variable across patients, ranging from 52.8% to 97.9%. Onset of a cluster was predicted with a mean F1-score of 74.6% (± 17.6%), with per-patient performance varying from 50.6% to > 90% (Table 2). Performance was better than naïve baseline classifiers for several patients in both tasks._x000D_
Conclusions: We demonstrate the feasibility of predicting seizure clusters based on ictal and pre-ictal iEEG data. The prediction is clinically important for managing seizure clusters in patients and can be valuable for exploring ways to mitigate their clinical burden._x000D_
Funding: This research was supported in part by NSF award CNS-1624790 and NIH award R01-NS92882.