Doctoral dissertation
Identification of hearth sounds for the analysis of cardiac pathology using machine learning
Views: 6 | Downloads: 7
Abstract
Cardiovascular diseases (CVDs) are a leading cause of mortality globally, significantly
affecting patient quality of life and imposing considerable demands on healthcare systems.
Chronic heart failure (CHF), a common outcome of CVDs, represents a growing health
burden with an increasing incidence. Accurate and early identification of CVDs is critical
for enabling timely diagnosis and effective treatment.
Artificial intelligence (AI) algorithms have shown promise in detecting CVDs, particularly
through the application of deep-learning methods using phonocardiogram (PCG)
data. However, these models require large datasets, which are often difficult to obtain.
Data augmentation has emerged as a viable solution to this challenge. This thesis investigates
advanced methods for improving the detection and classification of heart conditions,
focusing on CHF, and the development of a heart-sound-specific data-augmentation
method to enhance PCG data analysis.
The first investigation analyses heart sounds from CHF patients in their decompensated
and recompensated phases. The decompensated phase occurs when a CHF patient
requires hospitalization due to worsening symptoms, whereas the recompensated phase indicates
a patient’s subsequent stabilization. Using a PCG dataset from 37 CHF patients
and a combination of machine learning approaches, we achieved up to 72% classification
accuracy between the two phases, surpassing the cardiologists’ average accuracy of 50%.
Key predictive features were derived from diastole and included both time and frequency
domains.
The second investigation introduces PCGmix, a novel data-augmentation technique
specifically designed for PCG data. PCGmix generates new instances from original recordings
through meticulous interpolation, preserving key diagnostic features essential for CVD
detection. In this way, it also differs from existing data-augmentation methods, which
are general and not specialised for heart sounds. Empirical assessments using a publicly
available database of 851 normal and abnormal heart-sound recordings demonstrated that
PCGmix outperforms state-of-the-art augmentation techniques when data availability is
limited. Specifically, at 10, 30, 56, 112, and 168 PCG recordings, our method achieves the
same performance as the no-augmentation approach when train on the amount of data
that is 1.35–1.46, 1.34–1.69, 1.46–1.69, 1.08–2.25, and 1.01–1.65 times bigger, respectively.
This thesis contributes to the knowledge at the intersection of AI and medicine, offering
novel insights and methodologies to advance cardiovascular health monitoring and
diagnosis. It demonstrates the performance achievable in detecting CHF decompensation
using PCGs alone, and introduces and thoroughly validates PCGmix, a novel heart-soundspecific
data-augmentation method.