Brief Review — Deep Learning Methods for Heart Sounds Classification: A Systematic Review

An Overview Paper for Heart Sound Classification

Sik-Ho Tsang
5 min readNov 29, 2023

Deep Learning Methods for Heart Sounds Classification: A Systematic Review
CNN & RNN Overview, by Nantong University
2021 MDPI J. Entropy, Over 70 Citations (Sik-Ho Tsang @ Medium)

Heart Sound Classification
2013 [PASCAL] … 2022 [CirCor Dataset] [CNN-LSTM] [DsaNet] [Modified Xception] [Improved MFCC+Modified ResNet] 2023 [2LSTM+3FC, 3CONV+2FC] [NRC-Net]
==== My Other Paper Readings Are Also Over Here ====

  • In the field of heart sound classification, there are still limitations due to insufficient data, inefficient training, and the unavailability of effective models.
  • An in-depth systematic review and an analysis of existing deep learning methods were performed in the present study, particularly emphasizing on CNN and RNN over the last 5 years.

Outline

  1. Studies of Heart Sound Classification
  2. Heart Sound Classification Using Deep Learning
  3. Comparison of Deep Learning and Traditional Machine Learning Methods
  4. Limitations & Challenges

1. Studies of Heart Sound Classification

1.1. Preliminaries

PCG with simultaneous ECG recording

The above figure shows a PCG process with simultaneous electrocardiogram (ECG) recording and the 4 states of the PCG recording: S1, the systole, S2, and the diastole.

1.2. Previous Studies

Deep learning-based methods for heart sounds classification

Based on the databases of ScienceDirect, SpringerLink, IEEEXplore, andWeb of Science, 33 of the related studies are obtained as above.

  • To the best of authors’ knowledge, this is the first review report that consolidates the findings on deep learning technologies for heart sounds classification.

1.3. General Procedures of Automatic Heart Sound Classification

Automatic Heart Sound Classification
  • Automatic heart sounds classification process generally consists of 4 steps: denoising, segmentation, feature extraction, and classification.
  • Public heart sound databases rarely include synchronized ECG signals, which makes it difficult to segment the heart sound signals based on ECG signals.

2. Heart Sound Classification Using Deep Learning

The above figure shows a typical block diagram of a deep learning approach to heart sounds classification.

Heart Sound Classification Using Deep Learning
Heart Sound Classification Using Deep Learning

The approaches based on deep learning are mainly divided into CNN, RNN, and hybrid methods.

2.1. 2D CNN

  • e.g.: In authors’ previous study [33], they propose a method by using a combination of Inception and ResNet networks to develop a 138-layer CNN network.

Although 2D feature maps provide good representations of acoustically significant patterns, they require an extra transform procedure and the use of a set of hyper-parameters.

2.2. 1D CNN

  • Consequently, various 1D CNN-based methods with different CNN architectures have been proposed for identifying different kinds of heart sounds.

In a typical example, the 1D PCG time series is directly used as the 1D CNN without any spatial domain transform such as STFT or DWT.

  • e.g.: [37] proposed a novel 1D deep CNN for PCG patch classification.
  • 1D CNN obtains classification performance comparable with that of the 2D CNN without the requirement for feature engineering.

2.3. RNN

Heart sound signals are a kind of sequential data with a strong temporal correlation and can thus be suitably processed by RNNs.

  • e.g.: [48] were the first to use an RNN-based method, particularly using GRU, to detect anomalies in heart sounds provided by the 2016 PhysioNet/CinC Challenge.
  • e.g.: [46] examined 4 different types of RNNs; namely, LSTMs, BLSTMs, GRUs, and BiGRUs.

2.4. Hybrid

The most typical model-based integrated methods combine CNNs and RNNs. There are two reasons for this particular combination.

  • The first is that CNNs use various stacked convolution kernels to extract the features layer-by-layer.
  • The second reason is that RNNs process the signals of the timing relationship.
  • e.g.: [24] exploited the spatial and temporal characteristics extracted from the CNN and RNN, respectively, to achieve a higher accuracy.

There is also work which combines 1D CNN and 2D CNN.

  • e.g.: [52] The 1D CNN was designed to learn the time-domain features from raw PCG signals, while 2D CNN learned the time–frequency features.

3. Comparison of Deep Learning and Traditional Machine Learning Methods

Comparison of Deep Learning and Traditional Machine Learning Methods
Strength and Limitations of Deep Learning and Traditional Machine Learning Methods

3.1. Traditional Machine Learning

  • In most studies on traditional machine learning methods for heart sounds classification, a segmentation algorithm was used to identify the locations of the S1, S2, systole, and diastolic phases.
  • Also, traditional machine learning methods for heart sounds classification generally use small-scale training data, and the feature learning is based on prior knowledge of the data.

3.2. Deep Learning

  • Unlike traditional machine learning methods, the single architecture of a deep learning method can be used for joint feature extraction, feature selection, and classification.

4. Limitations & Challenges

  1. Limited heart sound data.
  2. Existing public heart sound datasets are usually scarce and imbalanced among different classes.
  3. Most concentrates on the two-class (normal and abnormal) problems of heart sounds classification.
  4. Diverse data distribution: Even the best model trained on the PhysioNet/CinC Challenge datasets [23] achieved only 50.25% accuracy when tested on the HSSDB dataset.
  5. Moreover, a more powerful deep learning models are needed, which normally exhibit more accurate performance. Yet, it needs higher computational power.

Therefore, the major problems requiring solutions include data insufficiency, training inefficiency, and insufficiently powerful models.

--

--

Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: https://linktr.ee/shtsang for Twitter, LinkedIn, etc.