Mobile phone-based rheumatic heart disease detection

Rheumatic heart disease (RHD), the permanent damage of the heart valves caused by an untreated “strep throat” infection, is the leading cause of cardiovascular mortality and morbidity in children and young adults worldwide. Simple penicillin treatment after the early diagnosis of RHD can stop recurring bouts of the condition, which lead to the most severe valvulopathy, and ultimately, heart failure. However, RHD is an under-diagnosed condition in the developing world, as such a diagnosis requires, at a minimum, a trained clinician to perform auscultation to detect pathological heart sounds. Trained medical personnel are scarce in the countries where RHD is most prevalent. A low-cost, mobile phone-based automatic diagnostic tool offers a potential solution, allowing a non-medically trained individual to screen for RHD in those countries. An essential feature of such a device is feedback on the signal quality of heart sound recordings. The first major contribution of this thesis is the investigation of features and algorithms for the automatic signal quality assessment of heart sound recordings. These algorithms are able to differentiate between good- and poor-quality recordings in over 80% of cases when using both a low-cost mobile phone-based stethoscope and an electronic stethoscope. Once the quality of recordings is ensured, the positions of the first and second heart sounds need to be located in a process called segmentation. This thesis extends the state-of-the art hidden semi-Markov models by: investigating additional features; extending the Viterbi algorithm; incorporating logistic regression into the model to form a hybrid generative-discriminative model; and investigating a discriminative duration-dependent probabilistic model – a conditional random field. These extensions are found to outperform the state-of-the-art method. Lastly, the period between the first and second heart sounds can be analysed for the presence of a pathological murmur. This thesis presents automated systolic murmur classification algorithms based on wavelet and mel-frequency cepstral coefficient-based features along with denoising via cycle averaging. These algorithms outperform three methods from the literature when detecting valvulopathy, while also outperforming a cardiologist and commercial software when detecting RHD in mobile phone-based heart sound recordings.