Recent advancements in signal processing, artificial intelligence, and data science have facilitated the exploration of computer-assisted auscultation as a frontline tool for cardiovascular disease detection. Despite its promise, the accuracy, effectiveness, and robustness of this approach require further evaluation. Particularly, while previous studies have shown encouraging results when training and testing on a single dataset, optimal ways have still to be found to combine varied data sources for training models that generalise well. This study addresses the critical need to evaluate the generalisability of heart sound classification models across different datasets, considering variations in data collection devices and acoustic environments. To the best of our knowledge, no prior study has assessed the generalisability of heart sound analysis methods across multiple datasets. Furthermore, many other studies solely report results on private datasets, hindering reassessment, reevaluation, and comparison efforts.
The objective of this study is to investigate the robustness of machine learning-based abnormal heart sound detection algorithms across multiple open datasets. We evaluate four publicly available and frequently used heart sound datasets, namely CirCor DigiScope Dataset, Heart Sounds Shenzhen, Pascal Challenge Database, and PhysioNet 2016 Challenge dataset. Our analysis focuses on the comparison of joint training with multiple datasets (also known as data agglomeration) and the late fusion of models trained on individual datasets (also known as majority voting). Through extensive experiments, we demonstrate the potential of combining multiple datasets to enhance cross-corpus heart murmur detection. Additionally, we explore data augmentation methods for further enhancing heart murmur detection under multi-dataset settings. This research underscores the importance of evaluating the generalisability of heart sound classification models across diverse datasets. By leveraging multiple datasets and employing advanced machine learning techniques, we advance the understanding of computer-assisted auscultation and pave the way for the development of more robust and scalable solutions for cardiovascular healthcare.