Diabetes Prediction from Clinical Data Using Deep Convolution Neural Network

According to the World Health Organization (WHO), diabetes is a rapidly growing chronic disease and the ninth leading cause of death across the globe. Thus, early diagnosis of diabetes is essential so that proper preventive action can be taken to avoid progressive effects. Diabetes is caused when the pancreas does not produce adequate insulin or when the human body cannot use the insulin produced by the pancreas effectively. Diabetes is related to metabolism and may cause kidney failure, lower limb amputation, heart attack, blindness, chronic renal failure, antipathy, cirrhosis of the liver, diabetic foot syndrome, neuropathy, transient hyperglycemia, hyperthyroidism, encephalopathy, stroke, adrenal gland tumors, glucagonoma, and many other difficulties. Diabetes has shown rapid growth in the population of low- and middle-income developing countries compared with high-income countries because of a lack of proper diet, stress, obesity, lack of exercise, etc. [1-4].

A healthy diet, maintaining a healthy body weight, regular physical activity, and avoiding tobacco use can commonly help delay or prevent diabetes. Generally, diabetes is categorized into two variants: type 1 (T1D) and type 2 (T2D). T1D diabetes is caused by a deficiency in insulin formation that controls blood sugar. The various symptoms of type 1 are thirst, constant hunger, excessive excretion of urine, vision changes, weight loss, and fatigue [3-5]. T2D is the most common diabetes that occurs due to inefficient use of insulin among total diabetes patients; more than 95% of people suffer from T2D diabetes. It shows symptoms such as physical inactivity, low immunity, and excessive body weight. Traditionally, diabetes is diagnosed using testing of blood sugar (AIC test), fasting blood sugar test, and glucose tolerance test. Early diagnosis of diabetes is very crucial for avoiding the critical complications that can occur due to diabetes [6-9].

Artificial intelligence (AI) approaches can now be used to diagnose a variety of diseases, with deep neural networks (DNNs) achieving the most outstanding results in classification tasks [10]. DNNs have been utilized to diagnose a diversity of disorders in recent years. The pancreas functions normally and produces enough insulin in the absence of diabetes. When insulin links to receptors on the cell's surface, the glucose molecule's entry into the cell is also opened. In T1D, the pancreas eventually stops manufacturing insulin, causing glucose transport to cells to be disrupted [11].

This article presents deep learning (DL)-based early diabetes prediction based on a healthcare database. The key contributions of the article are summarized as follows:

• Data normalization using novel adaptive min-max normalization for standardization of data and minimization of the outliers and anomalies in clinical data.
• Development of a DCNN-based diabetes detection system based on clinical data.
• GAN-based data augmentation is used to minimize data scarcity and class imbalance problems.
• Evaluation of the effectiveness of the proposed system is based on accuracy, precision, recall, and F1 score.

This article is arranged as follows: Section II focuses on the recent trends in diabetes prediction using clinical data. Section III presents the dataset and methods employed in the diabetes detection system. Furthermore, Section IV provides a detailed explanation of the proposed DL-based diabetes detection system. Afterward, Section V presents experimental results and an analytical discussion. Lastly, Section VI offers a conclusion and provides future directions for upgrading the proposed scheme.

Prediction of normal and diabetic persons is challenging because of less variability in raw clinical data, missing values, and incomplete data. Therefore, selecting a proper clinical dataset is essential for efficient diabetes prediction. The work aims to offer a distinctive attribute for diabetes detection and assess the significance of automatic DL-based diabetes detection. The flow of suggested diabetes prediction is illustrated in Figure 1, which encompasses the training and testing phases.

1.png

Figure 1. Flow diagram of methodology

The proposed framework includes data preprocessing, such as normalization, cleaning, and data augmentation (during training). Further, a lightweight DCNN is employed to depict and classify the features into diabetes labels. The DCNN model helps to learn the correlation and connection patterns in various medical, physical, and behavioural attributes of the patient. The early detection of diabetes helps patients to take preventive measures to avoid potential hazards to their health in the future. The outcomes of the system are assessed using accuracy and various qualitative and quantitative measures.

3.1 Dataset

The data for experimentation is selected from the PID dataset, which has an overall of 768 samples, 500 of which are healthy and 268 are diabetes samples [27]. It encompasses eight statistical, clinical variables such as age, blood pressure (BP), glucose, body mass index (BMI), glucose, insulin, triceps skin fold thickness (TST), count of pregnancies, and diabetes pedigree function (DPF) that can distinguish diabetes and non-diabetic samples in the dataset. Table 1 lists the dataset variable types and ranges in detail.

Table 1. Details of the Indian PIMA dataset

2a.png

(a)

2b.png

(b)

2c.png

(c)

2d.png

(d)

Figure 2. Analysis of dataset a) Diastolic BP Residual Plot; b) BMI Residual Plot; c) Diastolic BP Line Fit Plot; d) BMI Line Fit

Three variables are used as the outcome: BP and BMI as the dependent variables, and glucose level as the independent variable in the dataset. In that case, 46.66% of the overall population can be classified as having a positive outcome. This means that individuals who experience both an increase in their blood pressure level and their glucose level are at a high risk of developing diabetes and being affected by its aftereffects. The diastolic BP and the BMI residual plot are shown in Figure 2(a) and Figure 2(b). Figure 2(c) illustrates the regression plot for the diastolic BP analysis of the BMI vs glucose level.

The dataset consists of uneven sample sizes, which leads to a class imbalance problem. The diabetic samples are augmented to 500 samples to equalize the dataset size. The augmented dataset using GAN is given in Table 2. We have created the same samples using GAN and traditional SMOTE to analyze the performance of the system. The dataset is split into a 70:30 ratio for training and testing purposes. The 70% data for training provides stability in training, and the 30% data for testing provides enough samples for validating the effectiveness of the trained model for unseen data.

Table 2. Dataset details

3.2 Pre-processing of data

Pre-processing removes incomplete records, normalizes the samples, and updates missing values. The data with any missing values is removed from the original dataset. The preprocessing stage applies the novel adaptive min-max normalization to minimize the outliers and standardize the data. The original data (D) is normalized using Eq. (1), where $D_{\text {norm}}$ denotes normalized clinical data, $D_{\min}$ depicts the minimum value of clinical attribute, $D_{\max}$ indicate the maximum value of clinical data, and $\epsilon$ indicate samll positive number to avoid infinity value when $D_{\min}=D_{\max}$.

$D_{\text {norm}}=\min \left(1, \max \left(0, \frac{D-D_{\min}}{D_{\max}-D_{\min}+\epsilon}\right)\right)$           (1)

During the testing phase, a test sample is tested over the trained network that generates the recognition results into a healthy or diabetes sample. The five-layered DCNN architecture helps to characterize the distinguishing attributes of the raw diabetes clinical data.

3.3 Data augmentation using GAN

The GAN model is used for clinical data augmentation, which uses the generator (Gen) and discriminator (Dis) network as given in Figure 3. The GenNet creates the synthetic data from the random noise vector and tries to reduce the error in the synthetic and training samples. The DisNet network identifies the real and fake data. The framework of the GANbased data augmentation is shown in Figure 4. Eq. (2) provides GAN behavior where $x$ denotes real data, $z$ indicates the noise variable, $y$ denotes the conditioning variable, $D(x \mid y)$ specifies DisGen prediction for real input at condition $y$, and $G(z \mid y)$ indicates the sample generated by GenNet using noise vector at condition $y$.

$\min _G \max _D E_{x \sim P_r}[D(x \mid y)]-E_{z \sim P_z}[D(G(z \mid y))]$            (2)

3.png

Figure 3. GAN model for data augmentation

The GenNet consists of a reshape layer that converts the noise vector into 1×8 dimensions. The transpose convolution (TransConv) layers increase the feature connectivity. The three TransConv layers use 128, 64, 32, and 1 filters at four layers. The TransConv layers are followed by batch normalization (BN) for standardizing neurons and rectified linear unit layer (ReLU) to improve the features' non-linearity. The DisNet consists of four Conv layers with 128, 64, 32, and 1 filters at each layer, followed by BN and ReLU. The Softmax classifier is utilized to classify real and fake samples.

The GAN model is trained using the Adam optimization algorithm, considering a batch size of 16, an initial learning rate of 0.001, a cross-entropy loss function, and 100 epochs.

3.4 DCNN model

The DCNN helps provide multilevel hierarchical feature depiction, spatial connectivity between the features, and better correlation and connectivity between the local and global clinical attributes [28-30]. The clinical data has variability in the feature values, leading to complex feature depiction using DCNN. Also, there is a need to develop a lightweight DCNN framework that requires lower trainable parameters and time and can be efficiently implemented on standalone resource-constrained devices. The DCNN comprises five layers of CNN that consist of five convolution layers (Conv), five Rectified Linear Unit Layer (ReLU), one fully connected layer (FC), and lastly Softmax classifier layer (Softmax) as shown in Figure 4.

The first layer in the DCNN includes two layers {Conv1(KernelSize-1×3, NumFilter-16, Stride-1, ZeroPadding-Yes)→ReLU1 (Stride-1)} which accepts the diabetes data with dimensions of (1×8) and creates the output feature map of (1×8×16). Zero padding preserves the original sizes of the diabetes data. The second layer encompasses {Conv2(KernelSize-1×3, NumFilter-32, Stride-1, ZeroPadding-Yes)→ReLU3 (Stride-1)} which results in an output feature map of (1×8×32). The third layer encompasses {Conv3(KernelSize-1×3, NumFilter-64, Stride-1, ZeroPadding-Yes)→ReLU3 (Stride-1)}. Further, the fourth layer is designed as {Conv4(KernelSize-1×3, NumFilter-128, Stride-1, ZeroPadding-Yes)→ReLU4 (Stride-1)} that provides an output feature map of (1×8×128). The fifth layer includes {Conv5(KernelSize-1×3, NumFilter-256, Stride-1, ZeroPadding-Yes)→ReLU5 (Stride-1)} t (1×8×32).

The Conv feature map $y(n)$ of 1-D diabetes data $D(n)$ and Conv filter $k(n)$ with size L is given in Eq. (3). Eq. (4) indicates the Conv feature that describes the hierarchical connectivity features where $y_i^1$ denotes the $i_m^{\text {th }}$ feature map of $l^{\text {th }}$ layer, ${y}_{{j}}^{l-1}$ stands for the $\mathrm{j}_{\mathrm{m}}^{\text {th }}$ feature of $(l-1)^{\text {th }}$ layer, $\mathrm{k}_{\mathrm{ij}}^1$ denotes the filter kernel of $l^{\text {th }}$ layer connected to $j$ feature, $\mathrm{b}_{\mathrm{i}}^{\mathrm{l}}$ signifies for bias, and $\sigma$ denotes the ReLU function. The ReLU activation function is faster and easier to substitute negative neurons by 0 to overcome the vanishing gradient problem as given in Eq. (5).

$y(n)=D(n) \times k(n)=\sum_{m=0}^{L-1} s=D(m) \cdot k(n-m)$             (3)

$y_i^l=\sigma\left(b_i^l+\sum_j y_j^{l-1} \times k_{i j}^l\right)$           (4)

$\sigma(y)=\max (0, y)$             (5)

After five CNN layers, two FC layers were used, which had 50 hidden layers. Lastly, the SMC offers the output probability, where the label of the output neuron with the maximum probability yields the prediction result in Eqs. (6)-(8), respectively.

$z_i=\sum_j h_j w_{j i}$             (6)

$p_i=\frac{\exp \left(z_i\right)}{\sum_{j=1}^n \exp \left(z_j\right)}$                 (7)

$\hat{y}=\arg _i^{\max } p_i$              (8)

Here, ${h}_{{j}}$ is the weight of the second-to-last layer and $w_{j i}$ signifies the weights of SMC and the second-to-last layer, $z_i$ is the input of the SMC layer, $p_i$ is the likelihood of the class label and $\hat{y}$ is the predicted label. The DCNN is trained using the mini- batch gradient descent algorithm (MBGD).

4.png

Figure 4. The framework of DCNN for diabetes detection

The algorithm for the proposed DCNN-based diabetes prediction can be given as follows:

Step 1: Network and Hyperparameter Initialization

                N: Number of layers

                k: number of convolution filter

                α: Learning rate

                b: Initial bias

                w: Initial weights

                Num_feat: Number of features

Step 2:  Data Pre-processing

                Data Normalization

                Data Cleaning

Step 3:  Data Splitting

                Train_feat: Training Features

                Train_Labels: Training Labels

                Test_feat: Testing Features

                Test_Labels: Testing Labels

Step 4:  Training phase

                Layers

{

Convolution Layer1 (Stride-1, padding-same,k-16)

ReLU Layer1

Convolution Layer2 (Stride-1, padding-same,k-32)

ReLU Layer2

Convolution Layer3 (Stride-1, padding-same,k-64)

ReLU Layer3

Convolution Layer4 (Stride-1, padding-same,k-128)

ReLU Layer4

Convolution Layer5 (Stride-1, padding-same,k-256)

ReLU Layer5

Fully Connected Layer

Softmax Classifier

}

Net=train_network(Train_feat,   Train_Labels,Layers)

Step 5: Testing phase

                Predicted_label=test_network(Test_feat,   Net)

Step 6: Performance Evaluation

confusion_matrix=confusion(Predicted_label,Test_Labels)

                performance_metrics(confusion_matrix)

{

                Accuracy

Recall

Precision

F1-score                }