Related Experiment Videos
Neural network assisted cardiac auscultation
1Faculty of Health Sciences, University of Sydney, Lidcombe, NSW, Australia.
Artificial Intelligence in Medicine
|February 1, 1995
Summary
This study explored using neural networks for heart sound classification. Efficient training of these neural networks for phonocardiographic analysis was achieved, leading to accurate normal/abnormal heart sound identification.
Area of Science:
- Cardiology
- Biomedical Engineering
- Artificial Intelligence
Background:
- Cardiac auscultation traditionally requires significant interpretive expertise.
- Phonocardiography (heart sound recording) analysis can be complex.
- Artificial intelligence offers potential for objective heart sound interpretation.
Purpose of the Study:
- To evaluate the efficacy of neural networks as phonocardiographic classifiers.
- To identify optimal neural network training parameters (topologies, gain, momentum) for heart sound analysis.
- To develop a prototype classifier for normal versus abnormal heart sounds.
Main Methods:
- Three-layer neural networks were trained using backpropagation.
- The heart sound amplitude envelope was the sole input feature.
- Investigated various network topologies, gain, and momentum factors for efficient training.
- Developed a prototype normal/abnormal heart sound classifier.
Main Results:
- Neural network convergence was less likely with highly similar heart sound classes.
- The prototype classifier achieved excellent accuracy in distinguishing normal from abnormal heart sounds.
- Sparse training data did not impede the classifier's performance.
Conclusions:
- Neural networks show viability for computer-assisted phonocardiographic classification.
- Training parameters significantly influence the efficiency of neural network models for heart sound analysis.
- Further development can lead to a comprehensive computer-assisted phonocardiographic system.