Hybrid feature vector extraction in unsupervised learning neural classifier

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Heart rate variability (HRV) analysis is crucial for diagnosing cardiovascular diseases.
• Unsupervised learning methods offer potential for automated disease classification from complex physiological signals.
• Existing feature extraction methods may not optimally represent HRV for neural network classification.

Purpose of the Study:

• To develop and evaluate a new feature extraction and selection method for unsupervised learning neural classifiers using HRV signals.
• To create an optimal feature set from multi-domain HRV parameters for enhanced classification performance.
• To assess the efficacy of the proposed method using an Adaptive Resonance Theory (ART2) neural network.

Main Methods:

• A novel, multi-domain feature vector was constructed using time, frequency, and time-frequency HRV parameters.
• Feature ranking was performed using a class separability measure to select the optimal feature subset.
• An unsupervised neural classifier based on the Grosberg Adaptive Resonance Theory (ART2) was employed.
• The method was tested on 62 coronary artery disease patients, divided into learning and verifying sets.

Main Results:

• The feature selection process identified optimal features yielding the best classification results.
• The unsupervised ART2 neural network achieved performance comparable to multilayer perceptron structures.
• The reduced feature space representation improved the efficiency and effectiveness of the neural classifier.

Conclusions:

• The proposed feature extraction and selection method is effective for unsupervised HRV analysis.
• The ART2 neural network demonstrates competitive performance for cardiovascular disease classification using selected HRV features.
• This approach offers a promising strategy for developing robust, data-driven diagnostic tools in cardiology.