Abstracts

Rationale: Current ambulatory EEG systems are too cumbersome for long-term use in daily life. Single-channel, wireless, wearable sensors offer a method for capturing EEG data over extended periods of time in real-world settings. However, the manual review of long-duration EEG recordings is impractical, necessitating automated seizure detection. In this work, we present a machine learning-based approach that identifies common seizure types in a diverse patient population using single wearable sensors.

Methods: We developed a two-stage algorithm to detect three broad electrographic seizure types: tonic-clonic (TC), generalized absence (GA), and focal seizures (F). Scalp EEG was collected from 71 patients using a standard 19+ channel video-EEG system and single-channel wireless wearable REMI sensors (Epitel, Inc., Salt Lake City, UT), where 39 had TC seizures, 8 had GA seizures, and 24 had FIA seizures. Electrographic seizure start and stop times were determined by an epileptologist during video-EEG review. The raw, single-channel EEG data from the REMI sensors was preprocessed to account for noise and artifacts and then adaptively standardized and partitioned into 2-second non-overlapping segments. Twelve different features were extracted in the time and frequency domains out of which six were selected using mutual information and correlation analysis. All features were normalized using an estimate of their autocorrelation matrix. In stage I, separate ML classifiers (XGBoost) were trained to detect EEG segments during TC and GA seizures. The remaining segments were analyzed by a stage-2 classifier trained to detect segments during FIA seizures. A series of morphological processes were used to convert segment-level results into seizure events with estimated start and stop times.


Results: The algorithm was evaluated using 17-fold cross-validation (number of patients/fold: range = 3-4, median = 3), ensuring no subject's data appeared in both training and test sets. The algorithm was evaluated on 435 seizures (within fold range = 8-119, median = 17) recorded over 8,588 hours (within fold range = 246-863 hrs, median = 543 hrs), achieving an overall sensitivity of 78 ± 5.5% (SEM), and a false alert rate (FAR) of 1.95 ± 0.17/hr. For TC, GA, and FIA seizures individually, the sensitivities were 90 ± 2.8% (120 seizures), 75 ± 20.4% (153 seizures), and 66 ± 8.8% (162 seizures), with FARs of 0.07 ± 0.02/hr, 1.1 ± 0.36/hr, and 1.92 ± 0.34/hr, respectively.

Conclusions: We present a two-stage ML framework to detect common types of seizures using a single wearable EEG sensor. Results presented here suggest the potential to enhance support systems used by epileptologists for post-hoc reviews. Consequently, our algorithm represents an early demonstration of the feasibility of continuous, long-duration seizure monitoring using wearable EEG sensors during activities of daily life.


Funding: This work is supported by NIH NINDS grant U44NS121562