Development of novel optical sensing-based digital stethoscopes and deep learning-based heart sound denoising algorithms

This thesis describes the design and evaluation of novel optical sensing-based digital stethoscopes and deep learning-based heart sound denoising algorithms. Cardiac auscultation is the act of listening non-invasively to the sounds of the heart. It provides insights into the mechanical activity of the heart, and the sounds produced by this activity can be recorded as a phonocardiogram (PCG) using a digital stethoscope. Cardiac auscultation using a digital stethoscope is an important method for diagnosis of cardiovascular diseases (CVDs), such as valvular heart diseases, heart failure, and congenital heart defects (CHD). Digitally recording heart sounds as PCGs also paves the way for automated or AI-assisted diagnosis of CVDs. A literature review is conducted, including the background of cardiac auscultation, digital stethoscopes, classical signal processing approaches to denoising heart sounds, deep learning approaches, and heart sound processing at the edge. New research into digital stethoscope design and heart sound denoising are highlighted, and the current limitations faced in these areas are also explored. Two novel optical sensing-based digital stethoscope designs developed in the embedded systems group are considered in this thesis: one based on a reflective stethoscope diaphragm, and the other using a beam-cutting design. The goal of these designs is to address the limitations of current commercially available digital stethoscopes. The reflective diaphragm device is presented in conjunction with a simulation framework which can be used to test the device without live patients. The device is tested to be capable of accurately capturing a PCG signal and improving on the challenges faced by traditional digital stethoscope designs. The beam-cutting design is considered as suitable for a future printed circuit board (PCB) implementation, and to this end a bespoke testing platform is presented which facilitates testing of multiple optical components for this device and tuning of important design parameters. A number of tests are performed to characterise the performance of the components, resulting in the best being selected, along with optimum values for the design parameters. A complete pipeline for the denoising of heart sounds using fully convolutional networks (FCNs) is proposed, along with a thorough investigation of its capabilities and robustness. Both the hardware and software systems developed in this thesis were designed under the guidance of its co-supervisor, a medical doctor with clinical experience. Finally, the proposed deep learning-based denoisers are implemented on relevant edge hardware, leveraging the hardware acceleration of a neural processing unit (NPU).

Grantor dc:publisher

University College Cork

Year dc:date.issued

2025

Author dc:creator

• Duggan, Declan

Advisors dc:contributor.advisor

• Popovici, Emanuel
• Factor, Andreea

• Digital stethoscope
• Heart sound denoising
• Auscultation
• Phonocardiogram
• Signal processing
• Deep learning
• U-Net
• 3D printing
• Cardiovascular diseases
• Fully convolutional network
• Neural processor unit
• Hardware acceleration
• Edge AI

Statement dc:rights

• © 2025, Declan Duggan.

Licence dc:rights.uri

• https://creativecommons.org/licenses/by-nc/4.0/

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10468/18894

OAI identifier oai:identifier

oai:cora.ucc.ie:10468/18894