Detection of epileptic convulsions from accelerometry signals through machine learning approach

A seizure detection system in the non-clinical environment would enable long-term monitoring and give better insights into the number of seizures and their characteristics. Moreover, an alarm at seizure onset is important for alerting the parents or care-givers so they could comfort the child and optionally give the treatment. Therefore, we developed a patient-independent automatic algorithm for registration and detection of (tonic-)clonic seizures based on four accelerometers attached to the wrists and ankles. The objective is to classify two second epochs as seizure or non-seizure epochs employing supervised learning techniques. Starting from 140 features found in similar publications, a filter method based on mutual information is applied to remove irrelevant and redundant features. A least-squares support vector machine classifier is used to distinguish seizure and non-seizure epochs based on the selected features. For seizures longer than 30 seconds, median sensitivity of 100%, false detection rate of 0.39 h^-1 and alarm delay of 15.2 s over all patients are reached.