Automated Multimodal Epileptic Seizure Detection Using EEG and ECG
Abstract number :
3.459
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2018
Submission ID :
555129
Source :
www.aesnet.org
Presentation date :
12/3/2018 1:55:12 PM
Published date :
Nov 5, 2018, 18:00 PM
Authors :
Kaat Vandecasteele, Catholic University of Leuven; Thomas De Cooman, Catholic University of Leuven; Ying Gu, Catholic University of Leuven; Steven Vandeput, Catholic University of Leuven; Evy Cleeren, University Hospitals Leuven; Kasper Claes, UCB Pharma;
Rationale: In order to obtain an objective seizure diary outside the hospital, an automated wearable seizure detection device is required. To detect non-convulsive seizures, electroencephalography (EEG) and electrocardiography (ECG) are interesting biomedical signals. The combination of these signals for seizure detection is underreported in literature.Instead of using the full scalp EEG, only behind-the-ear EEG channels are used here. This can be measured in a comfortable and concealable way outside the hospital with a wearable device. The ECG is of interest since it was previously shown that temporal lobe seizures are often accompanied with strong ictal heart rate (HR) increases. Also the ECG can be recorded with a wearable device outside the hospital.This research proposes a patient-specific automated seizure detection algorithm based on behind-the-ear EEG and HR. Methods: The proposed algorithm combined 4 behind-the-ear EEG channels (two behind each ear) with the heart rate, recorded with two electrodes placed supraclavicularly left and right. These signals were recorded with the hospital hardware. The algorithm consisted of two sequential steps. Firstly, a Support Vector Machine-based algorithm generated alarms based on the EEG data only. Secondly, on the data segments responsible for generating these alarms, a k-nearest neighbors-based algorithm, using EEG and ECG data, was run. A key element in this algorithm was the alignment in time of the ictal manifestation in EEG and ECG.This algorithm was evaluated on a dataset of consecutive patients recorded in the University Hospitals Leuven between with the following inclusion criteria (1) at least 2 seizures with EEG correlates (2) at least an average ictal HR increase of 30 beats per minute. The dataset consisted of 1897 hours of data originating from 18 patients including 80 focal impaired awareness seizures, arising from (fronto-) temporal (78) and frontal-parietal (2) lobe. Results: The results are shown in Table 1 for the two steps of the algorithm. Mean and median were calculated for the Sensitivity (Sens), False Positives per hour (FP/h) and Positive Predictive Value (PPV). Conclusions: A multimodal EEG/ECG-based seizure detection algorithm is proposed. This algorithm results in a higher mean PPV (0.31) than the EEG-based algorithm (0.13). The ECG has added value to the behind-the-ear EEG channels for seizure detection. Funding: SeizeIT is a project realized in collaboration with imec, supported by VLAIO and Innoviris. Project partners are KU Leuven, UCB Pharma, Byteflies and Pilipili.