A Lightweight Robust Approach for Automatic Heart Murmurs and Clinical Outcomes Classification from Phonocardiogram Recordings

Abstract

Cardiac auscultation provides an efficient and cost-effective way for cardiac disease pre-screening. The George B. Moody PhysioNet Challenge 2022 aimed to detect heart murmurs and clinical outcomes with heart sound recordings from multiple auscultation locations. Our team HearHeart proposed a lightweight convolutional neural network (CNN) to detect heart murmurs and a random forest model to classify clinical outcomes. 128 Melspectrogram features and wide features like the socio-demographic data and statistical features are extracted. Different techniques are employed to migrate the data imbalance and model the overfitting problem. We used two data augmentation methods, noise injection and spectrogram augmentation in time and frequency domain to increase the training samples and avoid overfitting during training. Besides, weighted loss functions are applied to both tasks to deal with data imbalance. In the end, we ensembled the models from cross-validation and used voting for the final classification. We achieved a murmur score of 0.791, and a clinical outcome score of 11731.64 on 5-fold cross-validation in the hidden validation set. While on the hidden test set, we achieved a murmur score of 0.780, and a clinical outcome score of 12110, placing our team 1st and 10th in the challenge tasks, respectively.