EEG Feature Selection Techniques for Epileptic Seizure Detection: Performance and Evaluation Study

Abstract

Epilepsy poses a significant challenge for people around the world, particularly in third-world countries. Hence, over 80 million people are suffering from this disease. Therefore, the detection of epileptic seizures plays a vital role in diagnosis by measuring the activity of the brain, like the Electroencephalography (EEG) instrument. Nowadays, Data variety, attributes and size are growing increasingly, causing high dimensions and high preprocessing, which requires high computational resources. Therefore, reduce the data dimension to eliminate posing high computational resources in data preprocessing in the classification techniques. To address these issues, different feature selection techniques are used for high-dimensional data reduction by selecting the most relevant features in the classification process, which aim to detect disease faster and accurately. These techniques include Mean Decrease in Impurity (MDI), Correlation Coefficient, Sequential Forward Selection (SFS), and Sequential Backwards Selection (SBS). The results have registered diverse classification accuracy ratios for the mentioned method when using random forest classification, where MDI has 98.313% to 98.1% accuracy ranges for the selected features, whilst it achieved 98.504% in the case of using all the features. Moreover, the highest percentage, 98.4% accuracy, was achieved with the correlation method when extracting two features. On the other hand, SFS has an accuracy range from 98.6% to 98.212% after extracting nine features. A satisfactory classification accuracy was maintained by the SBS method, with accuracy from 98.6% to 98.2% after deleting ten features. Classification performance results show their impact in reducing dataset dimension, complexity, and computational resources cost, respectively.