Towards uncertainty-aware heart sound murmur detection via DirichletNet and tandem learning

Erika Bondareva, Jing Han, Tong Xia, Cecilia Mascolo
University of Cambridge


Abstract

Automated cardiac auscultation could promote preventative healthcare and improve the standard of care. In our work, instead of conventional three-class classification to predict murmur, normal, and unknown heart sounds, we propose to train two binary classifiers in a tandem fashion. Specifically, we first deploy a binary SVM classifier to distinguish unknown samples from all other samples. Subsequently, we extract large-scale handcrafted features and feed them into a deep neural network, trained to differentiate between murmur and normal samples. With this tandem learning strategy, we decompose the original complex task into two simpler ones and thus lower the risk of misclassifying the unknown samples. To validate our hypothesis, we performed patient-independent 10-fold cross-validation of our model on the training set, and achieved a score of 813. Furthermore, on the validation set, the algorithm scored 736, placing MobiHealth 32nd on the leaderboard.

While the present model is capable of undertaking the given task, it suffers from an unavoidable error propagation issue. If a misclassified unknown sample is passed through and classified as murmur/normal by the second classifier, the cost score would be heavily penalised. In reality, such an over-confident model would damage public trust in AI for health. To address this issue, we plan to combine the most advanced evidence-based uncertainty estimation approach, DirichletNet, with the present structure. In this way, the second module of our model could predict a Dirichlet distribution of the class probabilities for a given sample and thus quantify the model’s predictive uncertainty. This gives us an opportunity to mitigate the error propagation issue, by altering the prediction for high uncertainty samples back to unknown.