Optimisation and data mining techniques for the screening of epileptic patients

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Visual inspection of neurophysiologic signals like ElectroEncephaloGrams (EEGs) for abnormalities is difficult.
• Existing methods for analyzing complex, multi-dimensional time-series data often lack sufficient accuracy.

Purpose of the Study:

• To develop a novel classification technique for Multi-Dimensional Time Series (MDTS) data.
• To improve the accuracy and efficiency of differentiating between epileptic and normal subjects using EEG signals.

Main Methods:

• Proposed a novel Connectivity Support Vector Machine (C-SVM) technique integrating brain connectivity networks with Support Vector Machines (SVMs).
• Applied Independent Component Analysis (ICA) based on the Unbiased Quasi Newton Method to reduce noise in EEG data.

Main Results:

• C-SVM achieved 94.8% accuracy in classifying subjects.
• This represents a significant improvement compared to the 69.4% accuracy achieved by standard SVMs.
• The method demonstrated effectiveness in differentiating epileptic from normal subjects.

Conclusions:

• C-SVM offers a rapid and accurate method for online differentiation of epileptic and normal subjects.
• The proposed technique holds potential for solving other MDTS classification problems.