Brief Review — Abnormal heart sound detection using temporal quasi-periodic features and long short-term memory without segmentation


Sik-Ho Tsang
4 min readJan 21, 2024
AMDF for Feature Extraction, LSTM for Heart Sound Classification

Abnormal heart sound detection using temporal quasi-periodicfeatures and long short-term memory without segmentation
, by Harbin Institute of Technology, and Harbin Normal University
2019 Elsevier J. BSPC, Over 50 Citations (Sik-Ho Tsang @ Medium)

Heart Sound Classification
20132023 [2LSTM+3FC, 3CONV+2FC] [NRC-Net] [Log-MelSpectrum+Modified VGGNet] [CNN+BiGRU]
==== My Other Paper Readings Are Also Over Here ====

  • First, the spectrogram of the heart sound signal is extracted.
  • Then, the temporal quasi-periodic features of the heart sound signal are calculated.
  • Finally, long short-term memory (LSTM) is applied.


  2. Results


1.1. Average Magnitude Difference Function (AMDF)

  • In the first step, the heart sound signal is represented in the time–frequency domain by STFT. Suppose that the heart sound signal is denoted by x; therefore, the spectrogram of x, S, is extracted by:
  • Further, the temporal information can be analysed in different frequency bands of S. To improve the discrimination, the log power of S is calculated:
  • Subsequently, S is normalised to the value between 0 and 1.
  • To extract the temporal quasi-periodic features, the AMDF of each frequency band in the heart sound spectrogram S is calculated by:
  • Generally, 3 valleys exist in the AMDF of frequency f within the first average heart cycle, as shown above. In addition, similar valleys exist in other frequency bands.

It is found that the values of the abnormal heartsound AMDF features in the higher frequency bands are larger than those of the normal heart sound AMDF features. Thus, the AMDF features can distinguish between differentcategories of heart sound signals.

1.2. LSTM

  • AMDF features are fed into LSTM, which is composed of a series of LSTM units.
  • The extracted dependency relation features hT, by LSTM, are followed by a 2-layer neural network for classification. ReLU is used.
  • (After reading the paper, I think there is only 1 layer of LSTM.)
  • Because only two categories exist for classification, only one node appears in the output of the network. The loss function is defined by the mean square errors (MSE) and calculated by:
  • where ym is the label of the m-th heart sound sample in the testing dataset, and ˆym is the predicted value.
  • Yet, the dataset is imbalanced. The weighted loss function is defined:
  • With also threshold tuned instead of just using 0.5:
  • (It is a bit strange to me that authors use unweighted/weighted MSE loss instead of binary cross-entropy loss.)

2. Results

SOTA Comparisons on PhysioNet Dataset

The AMDF method using the temporal quasi-periodic features achieves an overall score 0.9008. Although this score is lower than the scores in [9] and [10], it is still higher than the scores in the methods of [13] and [14]. Therefore, the temporal quasi-periodic features are useful for abnormal heart sound detection.



Sik-Ho Tsang

PhD, Researcher. I share what I learn. :) Linktree: for Twitter, LinkedIn, etc.