Universum based Lagrangian twin bounded support vector machine to classify EEG signals.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Electroencephalogram (EEG) signals are crucial for detecting neurological disorders but are often noisy and contain outliers.
• Universum data, representing samples outside known classes, offers valuable insights into data distribution for improved classification.
• Existing methods like Universum Support Vector Machine (USVM) and Universum Twin Support Vector Machine (UTWSVM) have shown promise in EEG signal classification.

Purpose of the Study:

• To propose a novel method, universum based Lagrangian twin bounded support vector machine (ULTBSVM), for enhanced EEG signal classification.
• To leverage universum data to incorporate prior information about data distribution for classifying healthy and seizure EEG signals.
• To improve the stability and prevent overfitting in EEG signal classification models.

Main Methods:

• The ULTBSVM employs a strongly convex objective function using the square of the 2-norm of slack variables for unique solutions.
• It incorporates regularization terms aligned with the Structural Risk Minimization (SRM) principle, enhancing dual formulation stability.
• Interracial EEG data is utilized as universum data, and feature extraction techniques are applied to obtain noiseless features.

Main Results:

• The ULTBSVM demonstrated superior performance in classifying healthy and seizure EEG signals compared to USVM and UTWSVM across various EEG and UCI datasets.
• Analytical comparisons confirmed the effectiveness of ULTBSVM, particularly for real-world data, outperforming existing universum-based models and TWSVM in most cases.
• The method shows significant potential for accurate EEG signal classification.

Conclusions:

• The ULTBSVM is a robust and effective method for classifying EEG signals, including real-world datasets with interracial universum data.
• The study highlights the utility of universum data in improving classification accuracy for neurological disorder detection.
• The proposed framework can be extended for multi-class classification problems beyond binary classification.