[Paper Review] Deep CardioSound-An Ensembled Deep Learning Model for Heart Sound MultiLabelling

Guo Li, Steven Davenport+1 more|arXiv (Cornell University)|Apr 15, 2022

Phonocardiography and Auscultation Techniques33 citations

TL;DR

Proposes an ensemble deep learning model for multi-label annotation of heart sound recordings across labels such as murmur timing, pitch, grading, quality, and shape, achieving strong segment- and recording-level performance.

ABSTRACT

Heart sound diagnosis and classification play an essential role in detecting cardiovascular disorders, especially when the remote diagnosis becomes standard clinical practice. Most of the current work is designed for single category based heard sound classification tasks. To further extend the landscape of the automatic heart sound diagnosis landscape, this work proposes a deep multilabel learning model that can automatically annotate heart sound recordings with labels from different label groups, including murmur's timing, pitch, grading, quality, and shape. Our experiment results show that the proposed method has achieved outstanding performance on the holdout data for the multi-labelling task with sensitivity=0.990, specificity=0.999, F1=0.990 at the segments level, and an overall accuracy=0.969 at the patient's recording level.

Motivation & Objective

• Advance automatic heart sound diagnosis beyond single-label classification.
• Develop a multi-label framework covering diverse label groups (timing, pitch, grading, quality, shape).
• Demonstrate performance on holdout data with clinically relevant metrics.

Proposed method

• Use an ensemble deep learning architecture to perform multi-label heart sound labeling.
• Model designed to annotate labels from multiple groups including murmur timing, pitch, grading, quality, and shape.
• Train and evaluate on holdout data to demonstrate generalization across segments and patient recordings.

Experimental results

Research questions

• RQ1 Can a single model jointly predict multiple heart-sound label categories effectively?
• RQ2 What are the segment-level and recording-level performance achieved by the proposed ensemble for multi-label heart sound tasks?
• RQ3 How does the model perform across label groups such as timing, pitch, grading, quality, and shape?

Key findings

• Segment-level sensitivity: 0.990.
• Segment-level specificity: 0.999.
• Segment-level F1: 0.990.
• Patient recording-level overall accuracy: 0.969.