Automated Epileptic Seizure Detection Using Accelerometry, Heart Rate And Electromyogram
Abstract number :
1.354
Submission category :
18. Case Studies
Year :
2016
Submission ID :
194083
Source :
www.aesnet.org
Presentation date :
12/3/2016 12:00:00 AM
Published date :
Nov 21, 2016, 18:00 PM
Authors :
Thomas De Cooman, KU Leuven, Leuven, Belgium; Anouk Van de Vel, University of Antwerp; Berten Ceulemans, Antwerp University Hospital, University of Antwerp, Belgium; Lieven Lagae, University Hospitals KULeuven Belgium; and Sabine Van Huffel, KU Leuven, Le
Rationale: Around 30% of epilepsy patients cannot be cured by the use of anti-epileptic drugs. A possible solution to improve the quality of life of these patients is to use an automated warning system, which automatically detects ongoing seizures and warns the patient's caregivers when necessary. Such a system is very important for children with epilepsy and their parents, especially during the night. This research investigates which biomedical signals are the most beneficial for the automated detection of nocturnal tonic-clonic seizures in children in a home environment. Methods: Three different modalities are evaluated for their potential for automated tonic-clonic seizure detection: the electrocardiogram (ECG), electromyogram (EMG) and accelerometry (ACM). Three patient-independent seizure detection algorithms using only one of each modality are evaluated on a dataset containing over 252 hours of data from 7 children with 22 tonic-clonic seizures. The dataset contains one-lead ECG, 2 3D accelerometers (left wrist and right ankle) and 2 surface EMG sensors (one on each arm). Seizure annotations were made based on expert annotations using video-EEG monitoring. The unimodal algorithms are also combined with each other in order to investigate which modality combination leads to the best performance. Results: Table 1 gives an overview of the results for the unimodal algorithms and the combinations of the different algorithms. For the unimodal algorithms, the ECG algorithm was able to detect all seizures (100% sensitivity (Se)), but also resulted in 0.79 false positives (FP) per hour. Although the EMG and ACM algorithms resulted in a couple of missed seizures, they resulted in much less FPs (0.065 and 0.210FP/h). These missed seizures were typically caused by seizures with very short tonic or clonic phase ( < 5s), which did not cause a problem for the ECG algorithm. The EMG was able to detect the seizures the fastest with an average detection delay of 12.4s compared to the seizure onset. This is 1.8s faster than the ECG algorithm and 3.3s faster than the ACM algorithm. Adding the ECG algorithm to the EMG or ACM algorithm leads to a strong increase in performance, whereas adding the ACM algorithm to the EMG algorithm barely leads to a performance increase. This can be seen in the positive predictive value (PPV), which shows the percentage of correct alarms. Adding ACM to EMG leads to a small PPV decrease (but also a small Se increase), whereas adding ECG to EMG or ACM leads to strong PPV increases of 34.28% and 63.81%. Adding the EMG algorithm to the ECG+ACM combination only leads to 1 extra detected seizure and 1 less FP. Conclusions: Although the unimodal ECG-based tonic-clonic seizure detection algorithm performs less than the unimodal EMG- and ACM-based algorithms, the ECG is of great added value in multimodal approaches. The EMG-based algorithm showed to be the best unimodal algorithm, but ECG+ACM proved to be the best combination in case of performance and wearability. Funding: T. De Cooman is funded with an IWT PhD grant.
Case Studies