Brief Review — Supervised threshold-based heart sound classification algorithm

Extension of MFCC+CNN CinC’16 With Optimized λ

Sik-Ho Tsang
2 min readJan 20, 2024
Stethoscope (Free Image by Pixabay from

Supervised threshold-based heart sound classification algorithm
MFCC+CNN+Optimizied λ
, by Guangdong University of Technology, Guangdong Institute of Intelligent Manufacturing
2018 J. Physiol. Meas., Over 20 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 ====

  • MFCC is extracted, CNN in MFCC+CNN CinC’16 [19] is used.
  • λ is introduced and optimized by maximizing the prediction accuracy of the training instances.


  1. Proposed Method
  2. Results

1. Proposed Method

The heart sound instance is considered to be classified correctly, depending on the proportion of the total number of segments predicted accurately in M, and a threshold λ (0<λ<1).

  • Each abnormal and normal sample is divided into N^i_M (i = 1, 2, …, N1) segments and P^i_M (i = 1, 2, …, N2) segments.
  • If the i-th abnormal sample is predicted correctly, it means the value of
  • exceeds the λ.
  • Therefore, the value of:
  • exceeds (1-λ), indicating the i-th normal sample classified predicted correctly.

λ is continued tuning iteratively through gradient ascent/descent.

  • The iterations needed for seeking the optimal λ are 64, 317, 26, 121, 215, 112, 348, 65, 108 and 50 respectively.

2. Results

10-fold cross validation results
  • The detailed experimental results of cross-validation are given in Tab.2.
Confusion Matrix
  • The relevant confusion matrix of the results from the proposed algorithm is listed in Tab.3.
SOTA Comparisons

The proposed algorithm acquires a better MAcc.



Sik-Ho Tsang

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