Automatic analysis and classification of cardiac acoustic signals for long term monitoring

Abstract

Objective: Cardiovascular diseases are the leading cause of death worldwide resulting in over 17.9 million deaths each year. Most of these diseases are preventable and treatable, but their progression and outcomes are significantly more positive with early-stage diagnosis and proper disease management. Among the approaches available to assist with the task of early-stage diagnosis and management of cardiac conditions, automatic analysis of auscultatory recordings is one of the most promising ones, since it could be particularly suitable for ambulatory/wearable monitoring. Thus, proper investigation of abnormalities present in cardiac acoustic signals can provide vital clinical information to assist long term monitoring. Cardiac acoustic signals, however, are very susceptible to noise and artifacts, and their characteristics vary largely with the recording conditions which makes the analysis challenging. Additionally, there are challenges in the steps used for automatic analysis and classification of cardiac acoustic signals. Broadly, these steps are the segmentation, feature extraction and subsequent classification of recorded signals using selected features. This thesis presents approaches using novel features with the aim to assist the automatic early-stage detection of cardiovascular diseases with improved performance, using cardiac acoustic signals collected in real-world conditions. Methods: Cardiac auscultatory recordings were studied to identify potential features to help in the classification of recordings from subjects with and without cardiac diseases. The diseases considered in this study for the identification of the symptoms and characteristics are the valvular heart diseases due to stenosis and regurgitation, atrial fibrillation, and splitting of fundamental heart sounds leading to additional lub/dub sounds in the systole or diastole interval of a cardiac cycle. The localisation of cardiac sounds of interest was performed using an adaptive wavelet-based filtering in combination with the Shannon energy envelope and prior information of fundamental heart sounds. This is a prerequisite step for the feature extraction and subsequent classification of recordings, leading to a more precise diagnosis. Localised segments of S1 and S2 sounds, and artifacts, were used to extract a set of perceptual and statistical features using wavelet transform, homomorphic filtering, Hilbert transform and mel-scale filtering, which were then fed to train an ensemble classifier to interpret S1 and S2 sounds. Once sound peaks of interest were identified, features extracted from these peaks, together with the features used for the identification of S1 and S2 sounds, were used to develop an algorithm to classify recorded signals. Overall, 99 features were extracted and statistically analysed using neighborhood component analysis (NCA) to identify the features which showed the greatest ability in classifying recordings. Selected features were then fed to train an ensemble classifier to classify abnormal recordings, and hyperparameters were optimized to evaluate the performance of the trained classifier. Thus, a machine learning-based approach for the automatic identification and classification of S1 and S2, and normal and abnormal recordings, in real-world noisy recordings using a novel feature set is presented. The validity of the proposed algorithm was tested using acoustic signals recorded in real-world, non-controlled environments at four auscultation sites (aortic valve, tricuspid valve, mitral valve, and pulmonary valve), from the subjects with and without cardiac diseases; together with recordings from the three large public databases. The performance metrics of the methodology in relation to classification accuracy (CA), sensitivity (SE), precision (P+), and F1 score, were evaluated. Results: This thesis proposes four different algorithms to automatically classify fundamental heart sounds – S1 and S2; normal fundamental sounds and abnormal additional lub/dub sounds recordings; normal and abnormal recordings; and recordings with heart valve disorders, namely the mitral stenosis (MS), mitral regurgitation (MR), mitral valve prolapse (MVP), aortic stenosis (AS) and murmurs, using cardiac acoustic signals. The results obtained from these algorithms were as follows: • The algorithm to classify S1 and S2 sounds achieved an average SE of 91.59% and 89.78%, and F1 score of 90.65% and 89.42%, in classifying S1 and S2, respectively. 87 features were extracted and statistically studied to identify the top 14 features which showed the best capabilities in classifying S1 and S2, and artifacts. The analysis showed that the most relevant features were those extracted using Maximum Overlap Discrete Wavelet Transform (MODWT) and Hilbert transform. • The algorithm to classify normal fundamental heart sounds and abnormal additional lub/dub sounds in the systole or diastole intervals of a cardiac cycle, achieved an average SE of 89.15%, P+ of 89.71%, F1 of 89.41%, and CA of 95.11% using the test dataset from the PASCAL database. The top 10 features that achieved the highest weights in classifying these recordings were also identified. • Normal and abnormal classification of recordings using the proposed algorithm achieved a mean CA of 94.172%, and SE of 92.38%, in classifying recordings from the different databases. Among the top 10 acoustic features identified, the deterministic energy of the sound peaks of interest and the instantaneous frequency extracted using the Hilbert Huang-transform, achieved the highest weights. • The machine learning-based approach proposed to classify recordings of heart valve disorders (AS, MS, MR, and MVP) achieved an average CA of 98.26% and SE of 95.83%. 99 acoustic features were extracted and their abilities to differentiate these abnormalities were examined using weights obtained from the neighborhood component analysis (NCA). The top 10 features which showed the greatest abilities in classifying these abnormalities using recordings from the different databases were also identified. The achieved results demonstrate the ability of the algorithms to automatically identify and classify cardiac sounds. This work provides the basis for measurements of many useful clinical attributes of cardiac acoustic signals and can potentially help in monitoring the overall cardiac health for longer duration. The work presented in this thesis is the first-of-its-kind to validate the results using both, normal and pathological cardiac acoustic signals, recorded for a long continuous duration of 5 minutes at four different auscultation sites in non-controlled real-world conditions.