Brief Review — A heart sound classification method based on joint decision of extreme gradient boosting and deep neural network

XGBoost + LSTM

3 min readDec 5, 2023

A heart sound classification method based on joint decision of extreme gradient boosting and deep neural network
XGBoost + LSTM, by Shanghai Jiao Tong University, and Shanghai General Hospital
2021 J. Bio. Eng. (Sik-Ho Tsang @ Medium)
Heart Sound Classification
2013 … 2020 [1D-CNN] [WaveNet] 2023 [2LSTM+3FC, 3CONV+2FC] [NRC-Net]
==== My Other Paper Readings Are Also Over Here ====

First, the preprocessed heart sound recordings are segmented into four status, and 5 categories of features are extracted.
The first 4 categories of features are used as the input of the XGBoost classifier.
The last category is the Mel-frequency cepstral coefficient (MFCC), which is used as the input of long short-term memory network (LSTM).
Joint decision is made based on XGBoost and LSTM.

Outline

Feature Extraction
XGBoost & LSTM
Results

1. Feature Extraction

1.1. Dataset

PhysioNet dataset [6] is used with denoising and segmentation using Springer [5].
Then, feature extraction is performed.

1.2. Feature Set

Time, frequency, wavelet transform coefficient, and wavelet packet entropy features are extracted as above for XGBoost.
MFCC features are also extracted for LSTM.
In total, 136 features are extracted.

2. XGBoost & LSTM

2.1. XGBoost

Feature subtraction process is used to remove useless features.

2.2. LSTM

12 MFCC is extracted for each frame.
With 299 frames, 299×12 features are fed to LSTM.

2.3. Joint Decision

LSTM is used to classify the sound if it is normal, abnormal or unclassified.
If LSTM classifies the sound as abnormal or unclassified, XGBoost is used to classify the sound as normal or abnormal.