Automatic Non-linear Feature Selection Framework for Epileptic Seizure Detection

Abstract

This paper presents a qualitative automatic feature selection framework. Feature selection plays a very important role in selecting those features which provides the best results in terms of accuracy. The research work is aimed for in depth analysis of non-linear parameters using EEG signals. This paper also provides a comprehensive study of the features and their interpretations in characterizing epileptic seizures. We examine the quality of each feature independently in terms of classification performance metrics to provide meaningful information about the features. Optimally setting the non-linear combination leads to high classification accuracy. The accuracy can further improve by combining some other qualitative features with the optimal non-linear combination. Experimental results on two data sets shows that Hjorth parameters (HjPm) + approximate entropy (ApEn), and HjPm + ApEn + Higuchi fractal dimension (HFrDm) give high and nearly the same accuracy. However, HjPm + ApEn combined with some statistical feature gives more than 99% accuracy for most of the cases. The optimal combination of the features is providing a computationally inexpensive solution and can be deployed on low-cost hardware.