Brief Review — Automated heart sound classification system from unsegmented phonocardiogram (PCG) using deep neural network
F-NN Net-4 is Proposed
Automated heart sound classification system from unsegmented phonocardiogram (PCG) using deep neural network
F-NN Net-4, by Anna University, SRM Institute of Science and Technology
2020 J. Phys. Eng. Sci., Over 80 Citations (Sik-Ho Tsang @ Medium)Heart Sound Classification
2013 … 2021 [CardioXNet] [CNN & RNN Overview] 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 ====
- Deep neural network architectures such as a one-dimensional convolutional neural network (1D-CNN) and Feed-forward Neural Network (F-NN) are proposed for the classification of unsegmented phonocardiogram (PCG) signal.
- Finally, it is found that F-NN Net-4 is the best model.
Outline
- Dataset & Preprocessing
- Proposed Feed-forward Neural Network (F-NN)
- Results
1. Dataset & Preprocessing
1.1. Dataset
The problem formulation is the binary classification of normal and abnormal heart sounds.
This study uses Physionet Challenge 2016 dataset.
1.2. Preprocessing
The PCG records in the challenge 2016 dataset were recorded with a sampling frequency of 2000 Hz. The data records were down-sampled to a frequency of 500 Hz.
- Savitzky–Golay filter was used, which performs a data point smoothening to remove the noise component in the signal.
- The filtered PCG signal, which is divided into smaller time segments of 6s, is taken as the input layer.
2. Proposed Feed-forward Neural Network (F-NN)
- 4 network architectures are designed in this study.
3 networks Net-1, Net-2, and Net-3 use 1D-CNN; and Net-4, uses a Feedforward Neural Network (F-NN).
2.1. 1D-CNN
- 3 networks are designed with different configurations of the 1D-CNN layers.
- Net-1 uses one convolutional layer with a filter length of 100 and with 64 convolutional filters. Max pooling is used. It uses fully-connected layer at the head.
- Net-2 uses two convolutional layers with a filter length of 10 and uses 32 and 64 filters, respectively. Net-2 uses Global average pooling layer (GAP), as used in NIN, at the end before head.
- Net-3 uses one convolutional layer with a filter length of 5 and 32 filters. It also uses fully-connected layer at the head.
- All the convolutional layers use the rectified linear unit (ReLU) as the activation function.
2.2. F-NN
Net-4 consists only of MLP layer with 5 dense layers with the first layer with 1024 neurons with ReLU activation, second dense layer with 512 neurons and ReLU activation function, third layer with 128 neurons with ReLU activation, fourth layer with 64 neurons with sigmoid activation and fifth layer with 2 neurons and softmax activation layer.
3. Results
- The test data set is made up of 100 PCG records.
Net-4, which consists of only MLP layers, produced the highest sensitivity of 0.8673.
- Net-3 produced the highest detection of true negative cases with a specificity of 0.9298.
Net-4 provided the best overall performance compared to other DNN models with respect to MAcc of 0.8574. Besides, Net-4 resulted in the highest F1 score of 0.8458.
- Net-4 provided the highest Area under the Curve (AUC) value of 0.857.
Although Net-4 cannot obtain the highest accuracy among SOTA approaches, it has the least preprocessing steps. It does not need segmentation.
- The deep Feedforward neural net (Net-4) proposed in this study can detect abnormal heart sounds without much of pre-processing of the PCG signal. It can reduce the analysis time of screening the PCG records for heart disease identification, thus assisting the cardiologist in providing a faster treatment plan to the patients.
