Learning Front-end Filter-bank Parameters using Convolutional Neural Networks for Abnormal Heart Sound Detection.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Early heart disease diagnosis is crucial, especially in resource-limited areas.
• Current methods use Finite Impulse Response (FIR) band-pass filters with Convolutional Neural Networks (CNNs).
• State-of-the-art algorithms require fixed filter parameters, limiting adaptability.

Purpose of the Study:

• To develop a novel CNN architecture for automatic heart sound abnormality detection.
• To integrate learnable FIR band-pass filters within the CNN using time-convolution (tConv) layers.
• To investigate different initialization strategies and phase response constraints for the learnable filters.

Main Methods:

• Proposed a novel CNN architecture incorporating time-convolution (tConv) layers for learnable FIR filter-banks.
• Applied constraints to ensure linear and zero phase responses for the learnable FIR filters.
• Evaluated the model on the PhysioNet/CinC 2016 dataset using 4-fold cross-validation.

Main Results:

• The proposed CNN models outperformed the state-of-the-art system in heart sound abnormality detection.
• The linear phase FIR filter-bank method achieved a significant 9.54% absolute improvement in overall accuracy compared to the baseline.
• Different initialization strategies were explored for the learnable filters.

Conclusions:

• Integrating learnable FIR filters within CNNs offers a superior approach to automatic heart sound analysis.
• The proposed tConv layer method enhances the performance and adaptability of heart disease detection systems.
• This advancement holds promise for improving early diagnosis in various clinical settings.