Multimodal Approach for Epileptic Seizure Detection in Epilepsy Monitoring Units

Abstract

Epilepsy is one of the most common neurological disorders, affecting up to 1% of the World population. Patients with epilepsy may suffer from severe consequences from seizures (e.g. injuries) when not monitored. Automatic seizure detection systems could mitigate this problem, improving seizure tracking and alerting a caregiver during a seizure. Existing unimodal solutions for seizure detection, based on electroencephalogram (EEG) and electrocardiogram (ECG) still have an unacceptable level of false positives, which can be reduced by combining these two biosignals. In this paper, EEG and ECG data from 7 epileptic patients with diverse recording length and seizure types were used for analyzing the importance of multimodal seizure detection, at a total of around 110 h 2 m. A leave one seizure out cross validation was selected, grouping data containing the period before a seizure and the seizure period. A proof of concept of multimodal seizure detection which uses a deep learning architecture directly on raw data is performed - a Fully Convolutional Neural Network and an architecture based on LSTM were tested. The network based on LSTM achieved better performance - using the best of one or a combination of both signals, all patients had above 91% detected seizures, a specificity per epoch above 0.96 ± 0.06 and a detection delay below 8.5 ± 12 s. These results show potential for developing a patient-specific approach for seizure detection that can be transferred to the ambulatory.