Review — Automatic diagnosis of multiple cardiac diseases from PCG signals using convolutional neural network
5-Class Heart Sound Classification Using 1D-CNN
Automatic diagnosis of multiple cardiac diseases from PCG signals using convolutional neural network
1D-CNN, by Dr. A.P.J. Abdul Kalam Technical University, and Brno University of Technology
2019 J. CMBP, Over 30 Citations (Sik-Ho Tsang @ Medium)Heart Sound Classification
2013… 2019 [{AlexNet, VGG} + SVM] 2023 [2LSTM+3FC, 3CONV+2FC] [NRC-Net]
==== My Other Paper Readings Are Also Over Here ====
- Convolutional neural network (CNN) is proposed for heart sound classification.
- To improve the accuracy in a noisy environment and make the method robust, the proposed method has used data augmentation techniques for training and multi-classification of multiple cardiac diseases.
Outline
- Dataset & Data Augmentation
- Proposed Approach
- Results
1. Dataset & Data Augmentation
- Table 1: Original dataset in [27] contains 1000 audio samples, 5 classes, each class has 200 samples.
- Data standardization is applied after filtering.
- As actual heart sound recorded by the doctors can have some noise while recording the signal so dataset should have a noisy signal to perform better in real-life situations.
Therefore, the proposed method used background deformation technique as an audio augmentation. This deformation technique applied directly to the cardio signal before converting it into the input representation
2. Proposed Approach
After augmentation, preprocessing such as filtering is performed. Then the pre-processed signal is input to proposed CNN for classification.
2.1. Preprocessing
- Various noises and other artefacts are also recorded with these signals. Therefore, these noises and artefacts must be removed for the proper diagnosis of cardiac problems.
- Due to this, the amplitude of the signal may be varied to various ranges and its time length can also vary. Therefore, amplitude normalization is performed on all signals according to a 16-bit and signal time length has taken up to 2.5 sec.
- Recorded PCG signal usually contains also background noise, for example, high-frequency noise (voice, fan, etc.). The frequency of heart sound lies between 20 Hz — 150 Hz. Thus, higher frequencies greater than 150 Hz can be easily removed in the frequency domain.
- This approach used Gaussian Butterworth filter having low-cut at 150 Hz and high-cut at 20 Hz because Gaussian butter worth filter provides high-cut & low-cut as well as smooth filtering at the edges.
- After that resampling is done to convert this signal into 20000 vectors Figs 1 and 3.
2.2. CNN Architecture
In brief, 1D convolutional layers are used to extract deep features.
- ReLU and batch normalization are applied.
At the end, fully connected layer with Dropout are used.
- Cross entropy loss is used for model training.
3. Results
3.1. Evaluation
Both normal data training and with augmented data training methods achieved an average 96.23% and 98.60% accuracy respectively.
- Authors claim that more training data will provide greater accuracy, therefore in confusion matrix represent this same as greater accuracy with the larger dataset.
- (Indeed, the dataset size for evaluation is different. To me, it is difficult for comparison.)
The accuracy is high in the proposed model when the CNN model is used. Also, data augmentation improves the learning and accuracy of the CNN model.
- (Again, the dataset for evaluation is different for some of the approaches. To me, it is difficult for comparison.)
3.2. Model Integration into Low Computing Device
The average time taken by the proposed method is 0.506 sec in desktop and 9.169 sec in DSP processor.
