Supervised and Unsupervised Learning Systems as a Part of Hybrid Structures Applied in EGG Signals Classifiers

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Electrogastrography (EGG) signals are crucial for diagnosing gastric motility disorders.
• Analyzing non-stationary EGG signals presents challenges for traditional classification methods.
• Unsupervised learning offers potential for robust EGG signal analysis.

Purpose of the Study:

• To investigate unsupervised neural networks for EGG signal classification.
• To compare their performance against supervised perceptron networks.
• To evaluate a novel approach combining wavelet transform and self-organizing maps.

Main Methods:

• Feature extraction using wavelet transform to capture time-frequency characteristics of EGG signals.
• Application of self-organizing maps (Kohonen maps) for unsupervised classification.
• Testing on a dataset of 62 patients with various EGG rhythm disturbances and a control group.

Main Results:

• The proposed system effectively identified parameters in non-stationary EGG signals.
• Wavelet processing combined with Kohonen maps demonstrated high classification performance.
• Sensitivity and specificity exceeded 90% for the best classifier.

Conclusions:

• Unsupervised learning, particularly with wavelet transform and self-organizing maps, is a promising method for EGG signal analysis.
• This approach offers a robust alternative for diagnosing gastric motility disorders.
• The methodology shows potential for clinical application in EGG signal interpretation.