Background Information
According to World Health Organization [WHO] (2017), cardiovascular diseases (CVDs) are the number one cause of deaths globally. The reasons might be as follows. Firstly, many of the CVDs, such as coronary heart disease (CHD) and cardiomyopathy, are non-symptomatic in the early stage, making early detection and medical interference difficult. Secondly, over 75% of CVD deaths take place in low-income and middle-income countries (WHO, 2017). Patients with worse-off family background might have financial difficulty affording traditional CVD detection, like electrocardiogram (ECG) and echocardiogram. Thirdly, some elderly might mistake CVD symptoms for other trivial illnesses. Further, people who were not well-equipped with cardiovascular knowledge might overlook the symptoms or feel uncomfortable consulting a cardiovascular expert. There is a need to create an accurate, low-cost and convenient early CVD screening tool to alert people of their personal risk, which can lead to early intervention and disease prevention.
The proposed solution is to develop a cloud-based application which uses the built-in microphone of the mobile devices to record a user's heart sound, then use AI technology to denoise the data and finally generates a prediction on the user's risk of having CVDs. Heart sounds are sounds produced when blood flows across one of the heart valves that are loud enough to produce audible noise from the surface of the chest. Heart sound can be in normal pattern when recorded from a healthy individual or in an abnormal one, which may be a sign of a more serious heart condition or a structural defect in the heart itself, such as valvular heart disease or congenital heart disease. The basic idea of the research is using heart sound as indicators to conduct accurate and convenient screening of pre-symptomatic individuals. This idea is practical because the technology involves, built-in microphones without the need for additional hardware attachment. Here is a figure demonstrating the audio of my heartbeat recorded by the built-in microphones of an iPhone 7.
Many studies have already been conducted on theoretical level to classify heart sound, but practical translation is limited. The motivation of this research proposal is to make CVD early detection accessible to the public, especially for low-income household and the elderly.
Methodology
The methodology consists of four components: construction of an audio denoising system, construction of a murmur classifier, quality evaluation and construction of a cloud-based application.
The construction of an audio denoising system would be the main focus of the research. There are many existing approaches of audio noise reduction, but none of them target at performing noise reduction for an AI murmur classifier, which not only requires elimination of ambient noises but also the completeness of the heart murmurs. We will probably use some mathematical tools like Fourier Analysis, some statistical algorithm, like Kalman filtering which is the basis of a previous research paper (Salleh et al., 2012) to conduct noise reduction.
The classification of heart sounds is not a new topic. Many research studies have already been done which aim at designing an accurate and practical heart murmur classification systems. Various machine learning algorithms have been utilized in previous studies, including networks (NNs), support vector machines (SVMs), Fuzzy Neural Network with Structural Learning (FNNSL), just name a few (Dominguez-Morales et al., 2017). Among them there are two leading approaches from the PhysioNet/CinC Challenge 2016, which have reached the sensitivity of 94.24% and 86.91% as well as the specificity of 77.81% and 84.90% respectively (Dominguez-Morales et al., 2017). We might consider adapting their approach to build the heart sound classifier.
For the quality evaluation part, we have obtained a large quantity of worldwide public data and local in-house data. We will split the databases into two sets. 80% of the data will be selected random to become the training set, and the remaining will be left for evaluation use. This procedure will be repeated ten or more times to assure reliability.
Once we found the system is accurate and workable, we will start to make a cloud-based application for mobile devices which would be handy to use and free of charge. Firstly, the users need to download the app onto their mobile devices. Then, they would use the built-in microphones to record their heart sound and upload the recording. The cloud server will process the audio and return the results to the users’ device.
Reference
Dominguez-Morales, J., Jimenez-Fernandez, A., Dominguez-Morales, M., & Jimenez-Moreno, G. Deep Neural Networks for the Recognition and Classification of Heart Murmurs Using Neuromorphic Auditory Sensors. IEEE Transactions on Biomedical Circuits and Systems. 12(1), 24-34. DOI: 10.1109/tbcas.2017.2751545.
Salleh, S. H., Hussain, H. S., Swee, T. T., Ting, C. M., Noor, A. M., Pipatsart, S., Ali, J., & Yupapin, P. P. (2012). Acoustic cardiac signals analysis: a Kalman filter-based approach. International journal of nanomedicine, 7, 2873–2881. https://doi.org/10.2147/IJN.S32315
World Health Organization. (2017, May 17). Cardiovascular diseases (CVDs). https://www.who.int/news-room/fact-sheets/detail/cardiovascular-diseases-(cvds)
Please sign in
If you are a registered user on Laidlaw Scholars Network, please sign in