Classification of phonocardiogram signals to detect heart murmurs

Abstract

Phonocardiographs are used to capture high resolution recordings of heart sounds, including sub-audible sounds. Time series analysis of the recorded frequencies can be performed to uncover hidden structures in the data, and act as a decision support tool to aid in diagnosis of conditions such as heart murmurs. However, the complex structure and high-dimensionality of such data creates challenges - each recording contains around 40,000 frequency observations, with the exact number of observations varying between patients.

This talk will discuss the approach of analysing time series features in place of the raw data, and the use of such features in machine learning algorithms to classify phonocardiogram signals - identifying patients who do or do not have heart murmurs. Initial work has identified a subset of time series features which prove important inputs to classification algorithms, with preliminary results reporting accuracy of over 80%.