Ensembled Transfer Learning and Multiple Kernel Learning for Phonocardiogram Based Atherosclerotic Coronary Artery...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Conventional machine learning offers a basic approach for Coronary Artery Disease (CAD) detection via phonocardiogram (PCG).
• Integrating deep learning representations could enhance CAD detection performance metrics.

Purpose of the Study:

• To propose a novel Multiple Kernel Learning (MKL) method for fusing deep embeddings from Convolutional Neural Networks (CNNs).
• To improve the accuracy of non-invasive Coronary Artery Disease detection using phonocardiogram signals.

Main Methods:

• Utilized deep embeddings transferred from a pre-trained CNN.
• Developed a novel MKL approach to find optimal kernel combinations, maximizing similarity and minimizing redundancy.
• Conducted experiments on 960 PCG epochs from 40 CAD and 40 normal subjects.

Main Results:

• Transferred CNN embeddings achieved a maximum subject-level accuracy of 89.25% (kappa: 0.7850).
• Fusion of embeddings with handcrafted features using MKL resulted in 91.19% accuracy (kappa: 0.8238).

Conclusions:

• The proposed MKL fusion method demonstrates significant potential for developing high-accuracy Coronary Artery Disease detection systems.
• Phonocardiogram (PCG) signals offer a viable, non-invasive source for advanced CAD diagnosis.