Feature selection methods for accelerometry-based seizure detection in children.

We investigate the application of feature selection methods and their influence on distinguishing nocturnal motor seizures in epileptic children from normal nocturnal movements using accelerometry signals. We studied two feature selection methods applied one after the other to reduce the complexity and computation costs of least-squares support vector machine (LS-SVM) models. Simultaneous feature selection analyses were performed for each seizure type individually and jointly. Starting from 140 features, a filter method based on mutual information was applied to remove irrelevant and redundant features. The obtained subset was further reduced through a wrapper feature selection strategy using an LS-SVM classifier with both forward search and backward elimination. The discriminative power of each feature subset was evaluated on the test data in terms of the area under the receiver operating characteristic curve, sensitivity, and false detection rate per hour. We showed that, by using only a filter method for feature selection, it was possible to obtain classification results of comparable or slightly reduced performance with respect to the complete feature set. The attained results could facilitate further development of accelerometry-based seizure detection and alarm systems.