< Back to previous page

Project

Rheumatic Heart Disease Detection

Rheumatic heart disease (RHD) is a preventable yet serious public health
problem that affects millions of individuals in low- and middle-income countries,
as well as marginalized groups in developed countries. It is a potentiallyfatal cardiac disorder caused by damage to heart valves induced by one or
more episodes of rheumatic fever, an autoimmune inflammatory response to
group A streptococcal bacterium infection. RHD is most frequently diagnosed
in children. Effective early intervention can significantly reduce the risk of
premature death and disability associated with RHD. Once a patient has been
diagnosed with rheumatic fever, it is critical to avoid recurrent streptococcal
infections. Currently, a huge percentage of people with RHD go undiagnosed
or are only discovered after significant cardiac damage has already occurred.
Community-level screening programs could be a potential solution. However,
healthcare staffs are overburdened with other obligations in RHD-endemic
countries, and hence they do not routinely screen or discover cases. This thesis
presents a novel and easy-to-use machine learning-based approach to detecting
RHD in the community that can be used by a non-medical trained individual.
Phonocardiogram (PCG) and Electrocardiogram (ECG) recordings were used
during the study. Both PCG and ECG carry useful characteristics that aid in
the analysis of heart-related disorders due to RHD.
When the heart valves are affected by RHD, they leak or become stenosed,
resulting in turbulent blood flow in the heart chambers. This is called cardiac
murmur. The sensitivity of manual auscultation for cardiac murmur detection
was reported to be very low. This is mainly due to the inherent limitation
of the human auditory system in performing accurate cardiac auscultation.
This issue can be resolved by developing a machine learning algorithm for
automatic detection of murmur due to RHD. However, there were no publicly
available heart sound data from persons with RHD (PwRHD). Hence, the first
objective was to collect a novel heart sound dataset of PwRHD. Specifically,data composed of PCG and ECG from 170 subjects (124 PwRHD and 46 healthy
control (HC)) was collected in Ethiopia, one of the RHD endemic countries.This data was used to develop, train and test different classification models for
RHD detection and screening.
Heart sounds, in particular, are vulnerable to a variety of noise sources, including
environmental and physiological sounds. There are various methods for reducing
and removing noise from such signals. This thesis investigated the impact of
various denoising methods on the classification performance of a heart sound
binary classifier. When denoising was directly used, it does not improve the
performance but when applied indirectly, i.e. to improve the localization
of the fundamental heart sounds, a considerable classification performance
improvement was observed. Specifically, Wiener Estimation based spectral
subtraction had the highest impact. This study revealed that, when used
correctly, heart sound denoising can improve performance.
In clinical practice, ECG tests are used to confirm the diagnosis in PwRHD,
and the results are typically reported as minor symptoms. However, ECG is a
particularly desired instrument in the developing world due to its ease of use and
widespread availability. This thesis investigated the automatic delineation and
characterization of ECG recordings from PwRHD. PR intervals were elongated
in 47.2% of the recordings. The QRS duration was longer than the normal
range in 26.5% of the cases. The QTc interval was prolonged in 44.3% of the
cases. Bradycardia was found in 62.2% of the patients. This study indicated
that ECG can be a useful tool in the assessment of RHD cases.
The final challenge was to investigate the performance of different classification
algorithms in RHD detection and develop an RHD screening tool. Multiple
classic classification algorithms were compared on the RHD dataset. This studyexamined the performance of decision trees, linear discriminant analysis, logistic
regression, Gaussian Naive Bayes classifiers, SVMs, KNNs, ensemble bagged
trees, and CNNs. SVM had exhibited a better classification performance and
was selected to develop the RHD screening tool. Thirty-one features composed
of time, frequency, perceptual and acoustic domain were selected and used in
the development of the system. To evaluate the system’s performance, a nestedcross
validation approach was employed to mimic various RHD prevalence
rates in the population. For a prevalence rate of 5%, the algorithm achieved
an f1-score of 72.2± 0.8 %, a recall of 92.3 ±0.4 %, a precision of 59.2± 3.6
%, and a specificity of 94.8 ±0.6 %. This study demonstrated the possibility
of automatic auscultation for RHD screening in the general population. The
proposed technology can be used to undertake regular screening activities by
someone who is not medically qualified, freeing up medical practitioners to
focus on more critical clinical obligations.

Date:21 Sep 2017 →  28 Feb 2022
Keywords:ECG, Biosignal, Sensors, PCG, RHD
Disciplines:Sensors, biosensors and smart sensors, Other electrical and electronic engineering
Project type:PhD project