One step in the right direction: Making health screening more accessible through automated analysis of sound recordings of the heart.

Objective: With cardiovascular diseases (CVDs) posing a significant global health challenge, early detection becomes paramount. Our paper explores the potential of leveraging phonocardiograms (PCGs) to detect heart murmurs, a crucial indicator of CVDs, employing an innovative explainable multitask model.

Approach: Our two-stage multitask model integrates multiple instance learning (MIL), enabling us to predict murmurs' presence in single PCGs and derive sample-wise classifications with minimal annotations. Additionally, we fuse explainable hand-crafted features with a pooling-based artificial neural network (PANN), culminating in robust predictions of murmurs and clinical outcomes across multiple PCG recordings.

Main Results: Through rigorous qualitative and quantitative analyses, we demonstrate the efficacy of the MIL approach in generating valuable features and detecting murmurs across various time instances. Our model achieved remarkable performance metrics, including a weighted accuracy of 0.714 and an outcome cost of 13612 on the CirCor dataset.

Significance: This research marks a significant milestone, being the first to showcase the utility of MIL in PCG classification. Moreover, our model's explainability facilitates quantitative analysis, mitigating confirmation bias and enhancing trust in the predictive outcomes. Ultimately, our findings underscore the potential of MIL combined with handcrafted features in advancing both explainability and classification performance in healthcare applications.

Kudos to the dedicated team behind this pioneering work! Feel free to read the full paper for deeper insights into our approach.