Abstracts

Rationale:
A wearable automated seizure detection device could allow for long-term monitoring of patients with epilepsy and reduce seizure-related adverse events.
Method:
In this study, we have explored the use of a novel smart textile combining cardiac dynamics (electrocardiogram) and movement (3D accelerometry) to build a long-term monitoring solution for the detection of epileptic seizures. Participants were asked to wear a smart garment with non-invasive sensors (Hexoskin) during their stay at the Epilepsy Monitoring Unit of the University of Montreal Hospital Center. Video-EEG of recorded events were reviewed (blinded to data from the smart textile) and classified as “epileptic seizure”, “no seizure” and “disturbed state” (where the patient is physically or emotionally disturbed, for example crying, but not experiencing an epileptic seizure). R-R interval series were located from electrocardiographic data using the Pan and Tompkins QRS detection algorithm. A sliding observation window was used to extract features into the time and frequency domains simultaneously from both R-R interval series and accelerometry data. Unimodal and multimodal analyses were implemented using different configurations of Support Vector Machine (SVM) classifiers to detect epileptic patterns.
Results:
Twenty epileptic patients were included, but data from only 11 (who experienced a total of 34 seizures) were annotated and could be exploited for analysis. The ECG-based detector yielded a sensitivity of 65.52% and a positive predictive value of 67.54% while the sensitivity of the accelerometry-based detector reached respectively 76.65% and 72.12%. Interestingly, after combining cardiac and motion analysis, the performance was improved allowing to reach a sensitivity of 76.34% and a positive predictive value of 76.03%.
Conclusion:
Hexoskin smart textile is a promising solution for monitoring epileptic patients. The combination of cardiac dynamics and movements improves the classification scores as it provides a wider perspective of the seizure than the analysis of each modality independently. The developed detector needs to be validated on a larger dataset to allow for a more reliable assessment of seizure detection performances.
Funding:
:This study was funded by the TELUQ University, the Canada Research Chair on Biomedical Data Mining (950-231214), CIHR/NSERC, IVADO (EBA), and Canada Research Chair Program (DKN).