Abstract
Cardiovascular disease (CVD) is the leading cause of mortality, accounting for 30% of deaths worldwide. Early screening and real-time monitoring play a vital role in the detection and taking necessary action to reduce the risk of worsening heart disease. The initial suspicion often depends on the medical staff to listen to murmurs in the heart sound (recorded as phonocardiogram, PCG) during auscultation, a sign of deflection from the electrocardiogram (ECG) and other signs such as low oxygen saturation (SpO2) and changes in respiration. However, these screening methods rely heavily on the physician’s auscultation experience, the ability to interpret ECG signals, and the real-time monitoring of multiple physiological signals. In the last decade, the rapid development of wearable devices and machine learning, especially deep learning techniques, has enabled the miniaturised and portable CVD screening devices in primary care.Nonetheless, the existing devices mainly provide single physiological signal measurement and cannot perform reliable assistive diagnoses, which limits the applicability of these devices. As a result, it is of great value to integrate multiple signal measurements with diagnostic capabilities on miniaturised screening devices in the future. For long-term vision, the research aims at designing a multi-sensor miniaturised device for cardiac investigation and monitoring, however, this thesis aims at (1) developing machine learning techniques to improve the computer-aided diagnosis of PCG and ECG; (2) assessing the feasibility by experiments, utilising biomedical signal processing to eliminate the need of physical respiration sensor. Specifically, the proposed objectives and outcomes are as follows:
1) To assess the feasibility of localised ECG signal acquisition and analyse its usability for PCG segmentation. We experimentally investigated the time property of ECG and PCG signals at auscultation sites and the effect of ECG inter-electrode distance. Results showed that ECG signal could be acquired stably at auscultation sites within a small area (5 cm), which provides a theoretical basis for designing miniaturised integrated ECG-PCG devices. Furthermore, the accuracy and robustness of PCG segmentation have always been important issues affecting PCG recognition. The obtained temporal relationships in this study will also make the device to perform reliable PCG segmentation using ECG signals.
2) To investigate the optimal use of deep learning input and propose a reliable algorithm for PCG classification. The conducted research aimed at optimising the information to improve the classification performance of deep learning. On recurrent neural networks (RNNs), the study analysed input length’s influence on classification accuracy using Melfrequency cepstral coefficients (MFCCs). The results indicated that an overly short signal length, such as one second, will weaken the network classification capability (reduce about 2-3% in accuracy). A comparative study was performed on deep convolutional neural networks (CNNs) to assess optimum time-frequency representations (TFRs) as input features for PCG classification. The results showed that continuous wavelet transform (CWT) and chirplet transform (CT) were slightly better than other TFRs including short-time Fourier transfer (STFT), Wigner-Ville distribution (WVD) and Choi-William distribution (CWD). Meanwhile, the appropriate increase of the CNN capacity and architecture optimisation can improve the performance, while the network architecture should not be overly complicated. Using prior knowledge and experience from the mentioned studies, A Hierarchical Multi-Scale Convolutional Network (HMS-Net) was proposed and won the first prize in the CinC/PhysioNet 2022 PCG classification challenge.
3) To design the deep learning algorithm for the detection of paroxysmal atrial fibrillation using single-lead ECG. A two-stage RNN network was proposed during the China Physiological Signal Challenge 2021 (CPSC 2021) which had satisfying performance and held the advantage of low computing load. It showed promising potential for terminal equipment such as the miniaturized ECG-PCG device.
4) To provide accurate respiratory rate while eliminating the need of physical respiration sensor. A study was conducted in the ECG-derived respiration (EDR) field to assess the feasibility of extracting EDR from the localised ECG at the auscultation sites by experiments. Results indicated that the ECG acquisition location barely affected the calculated respiratory rate accuracy. This proved the possibility of providing reliable respiratory rate from the ECG-PCG device without adding an extra sensor. In addition, the study was also conducted to compare the effect of using embroidered and gel electrodes on the extraction of EDR. Despite the slightly poorer performance of embroidered electrodes compared with gel electrodes, embroidered electrodes showed potentials in future low-cost applications. Stress test were also conducted for EDR by experiments. Current results indicated that the artefacts caused by body movement affected greatly on the EDR extraction. Reduce the swing of the ECG wires or using wireless ECG instead may be feasible solutions to improve the performance.
Date of Award | 1 May 2023 |
---|---|
Original language | English |
Awarding Institution |
|
Supervisor | Ernest Kamavuako (Supervisor) & Yansha Deng (Supervisor) |