RHEUMATIC HEART DISEASE DIAGNOSIS 

By: Lorna Mugambi

Background

Rheumatic Heart Disease (RHD) is a cardiovascular disease common in low and middle income countries due to inadequate infrastructure to treat Group A strep throat infections. RHD is caused by damage to heart valves due to inflammation and scarring caused by repeat rheumatic fever episodes. Rheumatic fever is caused by an autoimmune body response to the Group A strep throat infection caused by Group A streptococcal bacteria. It mostly affects children in developing countries where poverty is widespread, and the bacteria can easily spread. The mitral and aortic valves, found on the left side of the heart, are mostly affected by RHD.

It is important to detect cardiovascular diseases early so that management of the disease can begin. Echocardiographic screening for early detection of RHD has been proposed as a method to screen for RHD, but it is time consuming, costly and only a few people are skilled enough to reach a correct diagnosis.

In this research project we are presenting the use of machine learning as a tool to analyze echocardiograms which will automate the screening process of diagnosing RHD. The World Heart Federation (WHF) has come up with guidelines/ criteria for subclinical disease for RHD in asymptomatic populations. The criteria include valve thickness as a morphological criterion. Another criterion used is the velocity of the mitral valve’s regurgitant jet and length in at least two views, including the parasternal long axis view (PLAX).

Accomplishments

 So far, an echocardiographic view classifier has been developed. It is able to classify an echo with a PLAX view and a NOT PLAX view. A web application based on Dash has also been developed. This has been used by cardiologists to label data of echocardiograms we have, which will provide data to train and test a multiclass echocardiographic view classifier and a valve damage classifier. To address the challenge of limited labeled echocardiogram datasets, we have explored the use of unsupervised contrastive learning, a self-supervised learning method to try and detect and classify valvular pathology brought on by rheumatic heart disease.
Contrastive learning is a type of self-supervised learning technique that learns to distinguish between similar and dissimilar data points. It's a powerful method for learning useful representations from unlabeled data. The main idea behind contrastive learning is to learn embeddings such that similar data points are close to each other in the embedding space and dissimilar data points are far apart. This is achieved by defining a contrastive loss function that encourages this behavior.

Model Architecture

Uses a ResNet pretrained model as an encoder. The last fully connected layer is adapted to output 256-dimensional embeddings. However, there is an additional projection head after the base encoder. This projection head is a small neural network that maps the output of the base encoder to a feature vector. The purpose of this projection head is to encourage the model to learn more abstract representations of the images. The model is trained on pairs of images, where each pair consists of two different augmentations of the same image. The model is encouraged to output similar feature vectors for these two images. However, unlike in the supervised case, the model is not given any information about which images should have different feature vectors. It has to learn this from the data itself. This is done using the NTXent Loss function, which computes the loss based on how similar or different the feature vectors are.

Figure 1: A sample of images from the dataset before and after applying data augmentations

Figure 2: A representation of the model architecture used.

Figure 3: Resultant output embedding generated by the model visualized using dimensionality reduction techniques; Principal Component Analysis, t-SNE and UMAP and classified by echocardiographic views

Next Steps

Fine-tune the model to make sure that frames from the same video are accurately classified in the same cluster.

Publications