Feature Generation and Dimensionality Reduction using the Discrete Spectrum of the Schrödinger Operator for Epileptic...

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epilepsy is a severe neurological disorder, second only to stroke.
• Magnetoencephalography (MEG) is crucial for localizing the epileptogenic zone.
• Manual analysis of long MEG recordings for epileptic spikes is time-consuming and prone to errors.

Purpose of the Study:

• To develop effective machine learning algorithms for rapid and accurate epileptic spike detection from MEG data.
• To improve the clinical diagnosis of epilepsy.
• To introduce novel feature sets for MEG signal characterization.

Main Methods:

• Proposed new feature sets for MEG signals based on Semi-Classical Signal Analysis (SCSA).
• Utilized a Random Forest (RF) classifier.
• Employed 5-fold cross-validation on a balanced dataset of 3104 frames (100 samples per frame, 2-sample step).

Main Results:

• Achieved average sensitivity of 93.68% and specificity of 95.08%.
• SCSA method demonstrated effective characterization of peak-shaped signals.
• Reduced feature vector size while improving spike detection accuracy.

Conclusions:

• The proposed SCSA-based features enhance the accuracy and efficiency of epileptic spike detection in MEG.
• This machine learning approach offers a promising tool for improving epilepsy diagnosis.
• The method provides a more reliable and faster alternative to manual visual assessment.