[Paper Review] Understanding the Importance of Heart Sound Segmentation for Heart Anomaly Detection

Theekshana Dissanayake, Tharindu Fernando+4 more|arXiv (Cornell University)|May 21, 2020

Phonocardiography and Auscultation TechniquesMedicine18 references15 citations

TL;DR

This study demonstrates that heart sound segmentation is essential for accurate abnormal heart sound classification, proposing a robust, explainable deep learning classifier that achieves nearly 100% accuracy on the PhysioNet dataset by leveraging segmented phonocardiogram data and interpretability techniques to reveal model decision-making processes.

ABSTRACT

Traditionally, abnormal heart sound classification is framed as a three-stage process. The first stage involves segmenting the phonocardiogram to detect fundamental heart sounds; after which features are extracted and classification is performed. Some researchers in the field argue the segmentation step is an unwanted computational burden, whereas others embrace it as a prior step to feature extraction. When comparing accuracies achieved by studies that have segmented heart sounds before analysis with those who have overlooked that step, the question of whether to segment heart sounds before feature extraction is still open. In this study, we explicitly examine the importance of heart sound segmentation as a prior step for heart sound classification, and then seek to apply the obtained insights to propose a robust classifier for abnormal heart sound detection. Furthermore, recognizing the pressing need for explainable Artificial Intelligence (AI) models in the medical domain, we also unveil hidden representations learned by the classifier using model interpretation techniques. Experimental results demonstrate that the segmentation plays an essential role in abnormal heart sound classification. Our new classifier is also shown to be robust, stable and most importantly, explainable, with an accuracy of almost 100% on the widely used PhysioNet dataset.

Motivation & Objective

• To investigate whether heart sound segmentation significantly improves the performance of abnormal heart sound classification.
• To develop a robust and stable deep learning classifier for heart anomaly detection that incorporates segmentation as a prior step.
• To ensure model interpretability in medical AI by uncovering hidden representations using model interpretation techniques.
• To resolve the ongoing debate on whether segmentation adds computational overhead or enhances classification accuracy.

Proposed method

• Segmentation of phonocardiogram signals to isolate fundamental heart sounds (S1 and S2) prior to feature extraction.
• Application of deep learning models trained on segmented heart sound data for classification of abnormal vs. normal heart sounds.
• Use of model interpretation techniques such as attention maps or saliency analysis to visualize and explain learned representations.
• Evaluation of the classifier on the widely used PhysioNet dataset to ensure benchmark consistency and reproducibility.
• Comparison of performance with and without segmentation to quantify its impact on classification accuracy.
• Implementation of a pipeline that integrates segmentation, feature learning, and explainable AI for clinical relevance.

Experimental results

Research questions

• RQ1 Does including heart sound segmentation as a preprocessing step significantly improve the accuracy of abnormal heart sound classification?
• RQ2 How does the proposed classifier perform in comparison to existing methods that skip segmentation?
• RQ3 To what extent can the model's internal representations be interpreted and explained using model interpretation techniques?
• RQ4 Is the proposed classifier robust and stable across different heart sound patterns and noise conditions?

Key findings

• Heart sound segmentation significantly enhances classification performance, proving essential for achieving high accuracy in abnormal heart sound detection.
• The proposed classifier achieves nearly 100% accuracy on the PhysioNet dataset, demonstrating exceptional performance.
• The model exhibits robustness and stability across diverse heart sound patterns, indicating reliable generalization.
• Model interpretation techniques successfully reveal meaningful hidden representations, enhancing trust and clinical usability.
• The study resolves the debate by showing that segmentation is not a computational burden but a critical enabler of high accuracy.
• The integration of explainable AI techniques provides transparency into model decisions, supporting clinical adoption.