Deep Learning-Based Murmur Detection and Murmur Characteristics Classification from Phonocardiogram
No Thumbnail Available
Date
2025-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
CVD is the main cause of death worldwide. The World Heart Federation reported a 20.5 million
death from CVD in 2021, by 2030 this number is expected to rise to 23.6 million, and more than
75 percent of CVD death occur in low and middle income nations, as they have less access for
equitable health care services. Phonocardiograph (PCG) is an affordable and portable instrument
which can record, play back, and provide a visual display for heart sounds. Blood flow
that is turbulent may cause the cardiac valves to vibrate enough to produce murmurs, which
produces audible heart sounds and abnormal waveforms in the PCG. Most risk factors can be
significantly reduced by early diagnosis. Although, PCG signals are useful for detecting heart
murmurs it needs a trained medical professional for its interpretation. This study proposes a
hierarchical multi-scale convolutional neural network (HMS-Net) based automatic murmur detection
and murmur characteristics classification using the publicly available CirCor DigiScope
PCG dataset. Previous research studies mainly focus on classifying heart sounds as either murmur
vs. normal or detecting a limited type of valvular heart diseases (VHDs) but this study
extends it by designing a pipeline that first determines whether a murmur is present or absent;
not only these but also it advances the analysis by further classifying five key murmur characteristics:
timing, shape, pitch, grade, and quality, which is a significant step beyond traditional
binary murmur detection. For the training and evaluation, it adopts the HMS-Net model developed
by the winner of the George B. Moody PhysioNet Challenge, with a few important
adjustments. In the original work, the dataset included an ’unknown’ murmur class due to expert
uncertainty to label them as present or absent, so they used a quality assessment method
for this class. In this study, the ’unknown’ class is excluded, as it does not represent a valid
clinical category, so that this study focused on clinically meaningful labels. Additionally, a
separate prediction pipeline is designed to detect murmur and simultaneously predict the five
characteristics. The core model architecture and preprocessing steps remain unchanged. Each
task are trained and evaluated independently using metrics such as accuracy, weighted accuracy,
F1 score, AUROC, and AUPRC. The proposed model achieved an accuracy of 93.1%,
and F1 score of 86.9% for the murmur detection task. For the murmur characteristic tasks, the
model achieved accuracy scores between 74.3% and 81.5%. This goes a step beyond simply
detecting whether a murmur is present to identifying the key five murmur characteristics. Identifying
these characteristics is very important because it helps to determine the type of VHDs,
which makes the model’s output more clinically useful. These results are promising, considering the small data and class imbalance. These results suggest that the proposed system can
support early screening to improve patient referrals to specialized care with more detailed murmur
information of the patient, which could help to identify the underlying VHDs so that it
helps patients to get early diagnosis and timly treatment. It will be very helpful to reduce earlier
deaths due to VHDs in resource limited communities.
Description
Keywords
Valvular heart diseases, Murmur detection, Murmur characteristic, Phonocardiograph, Deep-Learning, Classification