Detection of epileptic seizures based on video and accelerometer recordings

Epilepsy is one of the most common neurological diseases, especially in children. And although the majority of patients can be treated through medication or surgery (70%-75% a significant group of patients cannot be treated. For this latter group of patients it is advisable to follow the evolution of the disease. This can be done through a long-term automatic monitoring, which gives an objective measure of the number of seizures that the patient has, for example during the night. On the other hand, there is a reduced social control overnight and the parents or caregivers can miss some seizures. In severe seizures, it is sometimes necessary, however, to avoid dangerous situations during or after the seizure (e.g. the danger of suffocation caused by vomiting or a position that obstructs breathing, or the risk of injury during violent movements), and to comfort the child after the seizure. The current gold standard for the monitoring of epilepsy makes use of the combination of video and EEG. The recording of EEG is however hampered by the electrodes that must be attached to the scalp. This makes it uncomfortable for the patient to sleep with, and in addition, the attachment of the electrodes is labor-intensive which makes long-term home monitoring not feasible. In order to make this possible, we examine in this thesis whether the use of sensors that are less intrusive, namely accelerometers attached with wrist and ankle straps, and video, may be used to detect epileptic seizures with a motor component. This thesis focuses on two types of seizures in children, namely hypermotor and myoclonic seizures. After a preprocessing of the data to create individual movement events, first the individual modalities are used for detecting seizures. For this purpose, state-ofthe- art methods are used which are already applied in other fields. For example, for the detection based on accelerometer data, the comparison is made between a classification based on support vector machines and an outlier detection. The video detection on its turn uses features derived in the neighborhood of spatio-temporal interest points to classify movement patterns. This method is for example already used to recognize real actions in movies. An additional difficulty in the seizure detection using video, is that in comparison with the literature, the setup in this thesis attempts to disrupt the normal sleep of the patient as little as possible. To this end no markers or other visual aids are used, and the patient may keep using sheets. In a last part of this thesis it is examined whether an integrated approach can yield a better result. Here, the video and accelerometer data are joined, both at feature and at classifier level. The ultimate goal of the research is to create algorithms that are useful in practice. This has led to a home monitoring system that is currently in a test phase, in which patients are monitored for one month each. Hopefully this thesis can therefore contribute to improve the monitoring and thus the quality of life of patients.