Generative Adversarial Network Based Improved Transfer Learning Models for Chronic Heart Failure Detection

Existing methods are used to compare the results with proposed methodology. These existing methods are described in this section.

2.1 K nearest neighbors

KNN is a very simple instance-based learning algorithm with wide application in the area of classification tasks. The algorithm classifies any data point based on most of its 'k' nearest neighbors from a training dataset [9]. KNN creates predictions using the entire training set rather than explicitly creating a model, in contrast to many other machine learning algorithms.

KNN works with following steps:

Distance Calculation: To classify a particular data point, the KNN algorithm determines how far apart it is from each other point in the training set. One standard distance metric that is in everyday use is Euclidean, and for two points, $X=(x 1, x {2, \ldots . ., x \mathrm{n})})$ and $Y=(y 1, y {2, \ldots . ., y \mathrm{n}})$ in n-dimensional space, it comes out to be:

$d(x, y)=\sum_{i=1}^n(x i-y i)^2$            (1)

Finding Nearest Neighbors: After computing the distances, KNN finds the 'k' points in the training set that are nearest to that data point. Here, 'k' is a user-defined constant and must be specified depending on the problem at hand.

Majority Voting: The data point is assigned by the algorithm to the class that its 'k' nearest neighbors share the most. If k=1, it simply gets the class of its closest neighbor. For larger 'k,' pick the class with the highest frequency among neighbors.

2.2 1D convolutional neural network

1DCNNs are versions of CNNs constructed to work with one-dimensional sequential input data, time series, audio signals, or any data for which the sequence of elements matters. Unlike traditional CNNs, which are applied to 2D data, 1D CNNs apply convolutional filters along only one dimension to extract local patterns across a sequence [10]. The 1D-CNN design is such that a few layers are convolutional for feature learning from input sequences; some have pooling, with fully connected layers after that; these could model temporal dependencies and patterns efficiently.

2.3 2D convolutional neural network

2-D CNN was created with the purpose of processing two-dimensional input, such as pictures. A 2-D CNN convolves these portions of the input by sliding filters across both height and width to return feature maps that emphasize relevant spatial features, such as Saad cropping, shapes, edges, and textures. In order to increase the network's computational efficiency while preserving as much information as feasible, pooling layers shrink these feature maps in the following stage. Fully connected layers at the back classify images, detect objects, and segment them. 2D CNNs have been a backbone for many state-of-the-art computer vision systems on account of their skill to pick up hierarchical representation from visual data [11].

2.4 Long short-term memory

LSTM is a specially designed RNN architecture that enables information to be forecast or learned over very long sequences by solving traditional RNN issues, essentially the vanishing gradient problem. This is through the memory cells within LSTMs and gating mechanisms: the input gate, the output gate, and the forget gate. All these control inflow, outflow, and internal flow that passes in every cell. In this line, LSTMs may maintain and update their cell state through which they might capture long-term dependencies in sequential information [12]. Therefore, areas of application for LSTMs include time series prediction, natural language processing, speech recognition, and many more such related processes that need understanding of the context over a long sequence.

2.5 Auto encoder

A deep neural network called an autoencoder efficiently encodes and decodes data via unsupervised feature learning. First, the AE compresses the data into a lower dimension’s latent space and then restores the data by decompression. The first step would be to use this unsupervised deep learning approach to learn an auto encoder as a compressed representation of input data for the identification of heart disease. Next, use this representation to categories or identify abnormalities. At the output layer, neural networks known as auto-encoders attempt to reassemble their input [13].

2.6 Data sources

In this work, the publicly available 2016 PhysioNet Challenge [14] heart sound dataset is considered. 3,240 heart sound recordings, ranging in length from 5 seconds to more than 2 minutes, are included in the collection. Both healthy and ill people were among the varied persons who made these recordings, and they did so in a variety of settings outside of clinical settings. Table 1 shows the dataset description of PhysioNet Challenge 2016 dataset.

Table 1. Dataset description

The proposed architecture aims to expand a GAN-based model to address class imbalance in heart sound datasets and subsequently apply feature extraction, proposed Dl and transfer learning techniques for heart sound classification. This GAN framework is trained adversarial to generate balanced heart sound data, mitigating class imbalance issues. Following data generation, Feature extraction techniques such as MFCC, chroma, and spectral contrast are applied to capture discriminative characteristics from both real and synthetic heart sound data. The extracted features are then utilized as input to ML models such as KNN, KNN+ Feature Fusion, KNN+ feature Fusion+ hypermeters, DL classification models such as 1D CNN,2D CNN, LSTM, hybrid CNN-autoencoder and Transfer Learning Models such as VGG16, MobileNet for accurate classification of heart sounds across different cardiac conditions. By fine-tuning these pre-trained models on the balanced heart sounds dataset, the models can adapt their learned representations to better capture the unique characteristics of heart sounds, thereby enhancing classification performance. Figure 2 depicts the suggested architecture, and the processes are covered in the section following.

4.1 Load dataset and preprocessing

This step involves acquiring a dataset containing PhysioNet heart signal recordings. The dataset is labelled with different classes of heart signals, such as normal and abnormal then Label Encoding is performed. Label encoding is a ML approach that converts category input into numerical values. During this process, each unique category or label in the categorical variable is allocated a distinct integer value. In our work the labels are abnormal and normal heart sound signals which are converted into 0 and 1 using label encoding. Resampling a signal involves changing its sampling rate, which is done to prepare data for analysis at a different frequency.

4.2 Proposed GAN model

A Step GAN is employed for data balancing by generating synthetic heart sound signals to augment underrepresented classes. In order to improve the dataset's balance and support the development of reliable models for heart sound classification, the Generator generates new heart sound signals while the Discriminator separates actual from synthetic signals. The GAN Model's operation [30] is shown in Figure 3.

2.png

Figure 2. Proposed system architecture

3.png

Figure 3. Working of GAN model

Below are the steps to generate synthetic data using GAN model:

i) Calculate Label Count (Data Imbalance): Dataset analysed to determine the number of .WAV files per class (label). Calculated the count of samples for each class to identify any data imbalance issues.

ii) Get Label with Less Number of Class: Identify the class (label) with the least number of .WAV files. This is the class that needs generation. Figure 4 shows labels and its count.

4.png

Figure 4. Label count

iii) Apply Inception Model to Generate .WAV File.

Using an Inception-based model for .WAV file generation entails creating a GAN architecture that uses an Inception-like structure as the generator. Design the generator network architecture, which will have an Inception-like topology. Inception models often use numerous parallel convolutional routes with varying filter sizes.

Random noise and latent spatial representation are accepted by the generator. Create a discriminator network design that can differentiate between created and real data. WAV files.

A CNN architecture created for audio classification serves as the discriminator. Train the GAN model with a real-world dataset.WAV files. During training, the generator attempts to provide realistic results.WAV files that can deceive the discriminator while it learns to tell the difference between real and created.WAV files.

Generator: The generator is designed to generate data with a specific shape. It starts with a sequential model and adds layers sequentially. The first is a dense layer of 12*1000 units with a LeakyReLU activation function, no bias. Batch normalization is applied just after this layer. Further reshaping of the data in this example is into a 3D shape: (1000, 12). The next step will be to add three convolutional 1D transpose layers, gradually reducing the number of filters from 256 to 128, then to 64, and finally to 32. Batch Normalisation and LeakyReLU activation come after each convolutional layer. The last convolutional layer uses tanh as the activation. Assertions are included to ensure that the output shapes are per dimension. The generator model is then returned. Figure 5 depicts generator model architecture diagram.

Discriminator: The discriminator network is crucial in adversarial learning, particularly in models like GANs. Its principal function is to identify between genuine and fabricated samples created by the generator network. The ` discriminator` function constructs the complete discriminator network. Building it starts with an input layer of a given, predetermined shape. In order to minimize the spatial dimensions of the resulting tensor, this entails applying numerous inception modules sequentially with optional residual connections based on the truthy or falsy value of the `use_residual` parameter. Following these inception modules, global average pooling is applied. Lastly, a dense layer with a sigmoid activation function is used for binary classification, which determines if a transaction is genuine or fraudulent. Majorly, the discriminator network's role is to efficiently learn distinguishing characteristics from the input data and provide relevant knowledge regarding the same to a generator network in adversarial training. Figure 6 depicts desciminator model architecture diagram.

5.png

Figure 5. Generator model architecture diagram

6.png

Figure 6. Discriminator model architecture diagram

The model uses independent modules, inception modules and residual connections, to enhance its learning capacity without losing its stability during training. Filter sizes in inception module are described in Table 2 while Summary table with filter sizes and parallel path of generator and descriminator are shown in Table 3.

Table 2. Filter sizes in inception module (with parallel path)

Table 3. Summary table with filter sizes and parallel path

4.3 Feature extraction

MFCC, Fast Fourier Transform, Wavelet, and Feature Fusion Technique are just a few of the feature extraction methods that have been used thus far on the heart signal data. By using the MFCC, chroma, and spectral contrast approaches to extract these qualities from audio, crucial information about the audio's spectral shape, pitch content, and timbral aspects can be obtained. These characteristics support the ability of deep learning and machine learning classifiers to identify patterns and provide predictions using the information gathered from the audio signals. MFCC, Chroma, STFT, and Spectral Contrast were selected due to their proven effectiveness in capturing the essential characteristics of audio signals relevant to PCG classification. MFCCs are well-suited for modeling the timbral texture of sound and are widely used in audio recognition tasks for their ability to mimic human auditory perception. Chroma features capture harmonic and tonal content, which helps in identifying pitch-related anomalies in heart sounds. Spectral Contrast emphasizes the difference between spectral peaks and valleys, aiding in the detection of subtle variations in signal energy. STFT provides a time-frequency representation that preserves both spectral and temporal information, essential for analyzing dynamic patterns in heart sounds. Compared to wavelet transforms, which are effective but more computationally intensive and complex to tune, the chosen features offer a balanced trade-off between interpretability, computational efficiency, and classification performance. MFCC of normal and abnormal signal shown in Figures 7(a) and (b) and Scaled MFCC of normal and abnormal signal given in Figures 8(a) and (b).

7a.png

(a)

7b.png

(b)

Figure 7. (a) MFCC of normal signal (b) MFCC of abnormal signal

8a.png

(a)

8b.png

(b)

Figure 8. (a) Scaled MFCC of normal signal (b) Scaled MFCC of abnormal signal

4.4 Feature reduction

T-SNE,also known as t-distributed stochastic neighbor embedding. It is primarily used for reducing the dimensions of the data so that it can be visualized. But unlike PCA, which is mainly used for feature reduction, it can also, to some extent, provide feature selection by mapping high-dimensional space data onto lower-dimensional space while preserving its integrity. Feature Distribution is depicted in Figure 9.

9.png

Figure 9. Features distribution

4.5 Algorithms

K-nearest neighbors (KNN) used for heart sound classification by first extracting relevant features from heart sound recordings, such as frequency-domain characteristics or time-domain parameters. These features are utilised to represent each instance of a heart sound in a multidimensional feature space. Throughout the training phase, KNN retains these feature vectors alongside their respective class labels. Following steps performed for classification with KNN: KNN (MFCC Features), KNN (MFCC Features with TSNE Feature Reduction, KNN (MFCC with Spectral Amplitude / FFT Features), KNN (Wavelet Based Features), KNN (Feature Fusion Technique), KNN (Hyper Parameters Tuning).

4.5.1 KNN with feature fusion+ hyper parameters tuning

The fusion of features in K-nearest Neighbors involves the combination of several feature sets in the model for enhanced performance. It provides better representation as it would be more prosperous and more comprehensive with information about the data. This is done by concatenating in one set the abundances from different feature sets, scaling accordingly, and optionally applying dimensionality reduction methods. These feature sets are then tuned with hyperparameters associated with the optimization of some main parameters of the KNN algorithm, that is, the number of neighbors. KNN has a hyperparameter, k, which represents the number of neighbors to consider. Use techniques random search to search for the optimal value of k. Evaluate the performance of KNN with different values of k using cross-validation. Select the value of k that yields the best performance on validation set. Combining feature fusion and optimizing KNN hyperparameters is done random searches with cross-validation because of its capability to leverage several data aspects while determining the best configuration for a KNN model.

4.5.2 Proposed hybrid CNN + autoencoder

It combines CNNs for feature extraction and autoencoders for unsupervised learning. Use a CNN to process raw audio data or spectrogram images and extract meaningful features. Connect the output of the CNN to an autoencoder architecture to learn a compressed representation of the features.

A hybrid model combining an autoencoder with a 1D convolutional network for a binary classification problem. First, an efficient format of the input data was learnt by successively applying an autoencoder in an unsupervised pre-training way. The learnt features are then applied to the classification issue by reshaping this learnt representation and feeding it into a 1D convolutional network.

In this model, there will be a Dropout layer wherein, during training, a portion of the input units will be zeroed at every update. Likely, the latter prevents overfitting so that the model does not become too biased toward particular features from the input.

Output Layer: The predicted probabilities for the binary classification challenge are obtained using a fully associated Dense layer with softmax activation. CNN-autoencoder model architecture shown in Figure 10.

10.png

Figure 10. CNN-autoencoder model architecture

4.5.3 Transfer learning algorithms:

Pre-trained models VGG16 and MobileNet are applied on both balanced and unbalanced heart sound datasets.

VGG16:

A deep convolutional neural network model called VGG16 was created specifically for image classification. The network consists of 16 layers of artificial neurones, each of which improves prediction accuracy by sequentially interpreting image data.

VGG 16 Network Architecture: The network consists of five convolutional blocks followed by a classification block. Architecture Diagram of VGG16 shown in Figure 11.

Each convolutional block contains:

Step 1: Uses Conv1D layers for 1-dimensional convolution suitable for audio data.

Step 2: Applies ReLU activation for non-linearity.

Step 3: Uses max pooling for downsampling.

Step 4: Includes dropout for regularization (preventing overfitting).

The classification block contains:

Step 1: Flattens the convolutional output into a 1D vector.

Step 2: Uses fully-connected layers (Dense) with ReLU activation for learning higher-level features.

Step 3: Outputs through a final dense layer with:

MobileNet:

Google developed MobileNet, a convolutional neural network (CNN) architecture, especially for embedded and mobile vision applications. The goal is to develop efficient, lightweight models that can be utilised on low-resource devices. To reduce the amount of parameters and computational complexity without sacrificing performance, the architecture places a high priority on the use of depthwise separable convolutions.

MobileNet Network Architecture:

Step 1: Input Layer: input_layer: This defines the input layer that receives the pre-processed WAV file data.

Step 2: Convolutional Blocks (1-5):

Each block follows a similar pattern:

DepthwiseConv1D: Applies a depthwise separable convolution to extract features efficiently.

Conv1D: Applies a pointwise convolution to increase the number of filters.

MaxPooling1D: Preserves significant properties while reducing the data's dimensionality.

Dropout: Randomly drops a percentage of activations to prevent overfitting.

Classification Block:

Flatten: Transforms the convolutional blocks' multi-dimensional output into a single dimension so that it may be fed into fully connected layers.

Dense layers (dense1-dense3): These layers, which are fully integrated, discover intricate connections between the heart sound classes and the retrieved attributes.

Output: The final output layer with the number of units depending on the output_number and activation function based on the problem_type. (Softmax for classification, linear for regression).Figure 12 depicts the mobilenet architecture diagram.

11.png

Figure 11. VGG 16 architecture diagram

12.png

Figure 12. MobileNet architecture diagram