Listening to the Heart: Unifying Open Audio Databases for Cardiology Research

While computer-assisted cardiac auscultation is highly promising, its accuracy, effectiveness, and robustness require rigorous evaluation.Particularly, to the best of our knowledge, no prior study has assessed the generalisability of heart sound analysis methods across multiple datasets.In this work we aggregated multiple open source datasets to study the robustness of machine learning-based abnormal heart sound detection algorithms.Specifically, we evaluated a machine learning model on four publicly available heart sound datasets under four different crossvalidation settings: within-corpus, cross-corpus, and two multi-corpus settings (data aggregation and decision aggregation).Our findings reveal that the multi-corpus setting with data aggregation outperforms the cross-corpus setting, suggesting that combining varied data sources enhances generalisability.However, despite this improvement, there are still challenges that require further investigation, which we discuss in detail.Overall, the study emphasises the need for clear protocols in data collection, labelling, and sharing to ensure fair comparisons and a deeper understanding of model generalisability.