Unsupervised Discovery of Echocardiographic Views for Rheumatic Heart Disease Diagnosis

Abstract

RHD is a cardiovascular disease that causes damage to the heart valves. If the damage is severe it is rectified using expensive valve replacement surgery. Early diagnosis of the disease allows for cost-friendly preventive measures. Specific views of the heart are required for proper assessment by heart specialists. Since routine screening is recommended for the rapid early identification of RHD, a large amount of patient data is generated. To handle this influx of data, trained AI is being used to automate view classification, unfortunately, the high cost of obtaining expert-labelled data in terms of time and money is prohibitive. Thus, we explore how the use of unsupervised AI methods can aid experts in the faster labelling of the data and what patterns PCA and agglomerative clustering identify in echo videos. We found that after appropriate preprocessing, these unsupervised methods can group videos with similar echocardiographic views. We also found that these methods were sensitive to the specific machines used to acquire the data and therefore care should be taken when applying them to data collected using different machines.