Abstracts

Rationale: Different biosensors have been used for seizure detection in patients with epilepsy. This study aims to assess the applicability of multiple biomedical signal combinations for seizure detection. Methods: During long term video EEG monitoring, we assessed photoplethysmography (PPG), electrodermal activity (EDA), EKG, temperature, and accelerometry (ACC) signals. Findings were correlated with seizures detected by EEG. A combination of six features extracted from these signals was adopted for seizure detection. A low pass filter was used to remove the high frequency noise in EDA and band pass filters with cut-off frequency of 0.4 and 30 Hz were used to preprocess ACC and PPG data. Three features were computed from time-domain and frequency-domain analyses of preprocessed ACC data. Two features were estimated from the EDA signal derivative, and the last feature is the 30 second moving average of heart rate extracted from time-frequency analysis of the PPG signal. The data was pre-processed and analyzed using Matlab2017a (MathWorks Inc.). Results: We prospectively enrolled 227 patients from the Epilepsy Monitoring Unit at Boston Children’s Hospital. In 31 patients, we captured 34 generalized tonic clonic seizures (GTCS) and 36 complex partial seizures (CPS). Patients’ age ranged from 5 months to 27 years (mean=13). For GTCS, results verify previous findings that ACC might be the most reliable signal for seizure onset detection, since these seizures are accompanied by body movements. The combined ACC feature was calculated and shown to be 100% indicative of seizure movements and with 100% accuracy can distinguish between random daily movements and movements generated from a GTCS.  Features from EDA provided additional information for early detection of GTCS and CPS. While PPG is very sensitive to motion artifacts, heart rate in general increases during and after GTCS. In GTCS the increase in heart rate was still evident in patients’ signal recordings 10 minutes after seizure onset when movements subsided, and therefore heart rate may serve as a metric for seizure verification in GTCS. However, heart rate changes following CPS onset are not consistent for all seizures, and some patients experienced a heart rate increase, while others experienced a decrease or no change. Conclusions: Features estimated from accelerometry signals can be used to detect GTC seizure onset. Heart rate generally increases during and after GTC episodes, and can be specifically used for verifying seizure episodes detected by ACC and EDA features. For CPS, EDA features show better performance for seizure detection than ACC and PPG features. Next steps may include validation through implementation of these algorithms in prospective cohorts. Funding: Thank you to the Epilepsy Research Fund for generously funding this project.