Evaluation of time domain features using best feature subsets based on mutual information for detecting epilepsy

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Epilepsy detection often relies on complex signal analysis.
• Previous studies have explored various features for electroencephalogram (EEG) analysis.
• Novel feature extraction methods are crucial for improving diagnostic accuracy.

Purpose of the Study:

• To investigate the efficacy of novel time-domain (TD) features extracted from discrete wavelet transform (DWT) for epilepsy detection.
• To compare the performance of these TD features with existing methods like mean absolute value (MAV).
• To evaluate the classification accuracy using different machine learning algorithms.

Main Methods:

• Extraction of TD features (waveform length, zero-crossings, slope sign changes) from DWT.
• Application of feature selection and ranking using mutual information (MI).
• Classification using Naive Bayes (NB) and Support Vector Machines (SVM) on University of Bonn and CMCH datasets.

Main Results:

• Naive Bayes (NB) achieved 100% classification accuracy (CA) distinguishing normal and epileptic data using the top 4 ranked TD features from DWT.
• NB also achieved 100% CA with the top 2 ranked features on clinical data from CMCH, India.
• The study demonstrated the effectiveness of novel TD features derived from DWT for epilepsy detection.

Conclusions:

• Novel time-domain features extracted via DWT show high potential for accurate epilepsy detection.
• Feature selection based on mutual information is effective in identifying significant features.
• The proposed method, particularly with Naive Bayes, offers a promising approach for real-time epilepsy diagnosis.