Deep intelligent predictive model for the identification of diabetes

1.   Introduction

Diabetes is a metabolic illness characterized by inadequate insulin production and secretion abnormalities ^[1], with hyperglycemia as the primary symptom. The physiological system will be harmed by long-term exposure of organs to hyperglycemia. Damage to tissues and organs, such as the eyes, kidneys, nerves, heart, and blood vessels, can result in organ and tissue failure. Diabetes mellitus is currently split into two types: diabetes mellitus type 1 and diabetes mellitus type 2. Type 2 diabetes is the most prevalent type of diabetes, accounting for approximately 95% of diabetic people ^[2,3]. The major causes of diabetes are environmental factors and poor lifestyle choices. Furthermore, diabetes is triggered by aging, poor diet, and inadequate activity ^[4].

The normal transition from health to diabetes mellitus occurs in three stages: health, pre-diabetes, and type 2 diabetes ^[5]. After being diagnosed with diabetes, patients' blood glucose levels will continue to rise, and pharmaceutical therapy is difficult to discontinue ^[6,7]. On the other hand, can maintain blood glucose control and even restore health with artificial intervention. Many studies have demonstrated that the most effective technique for avoiding and controlling diabetes is to detect and treat it as soon as possible. As a result, early detection and lifestyle adjustments are crucial in diabetes care and the avoidance of complications, including diabetic retinopathy ^[8,9,10,11].

Early diagnosis reduces the relative and absolute risk of heart attacks and mortality, as per the ADDITION-Europe analytical model analysis ^[12]. Diabetes mellitus is a global public health concern, with the International Diabetes Federation estimating that 463 million people worldwide have diabetes, with a 51% increase projected by 2045. Furthermore, it is predicted that for every diagnosed person with diabetes, there is one undiagnosed individual ^[13,14]. Understanding the evolving pattern of diabetes at all stages and extracting useful diabetes information from physical examination data is crucial to diabetes prevention and treatment. As a result, several scholars have tried to develop diabetes prediction models. Traditional statistical learning procedures, such as linear regression, were used to create the first models. Traditional statistical learning systems, on the other hand, provide predictions, but predictive accuracy is not one of their strong suits.

The toolbox was recently enhanced to cover a variety of machine-learning approaches. These approaches anticipate future occurrences based on patterns discovered in training data from previous examples. Furthermore, different techniques are sometimes combined to construct ensemble models that outperform single models in terms of prediction. For example, Kalsch et al. ^[14] established a relationship between liver damage markers and diabetes and used random forests to predict diabetes using serum data. Zou et al. ^[15] utilized principal component analysis (PCA) and minimal redundant maximum (mRMR) correlation to screen risk factors, and decision tree (DT) and Random forest (RF) to predict diabetes. However, they are based on traditional learning models, which need much human expertise and the ability to extract crucial parts automatically.

Furthermore, in the presence of nonlinearity in the dataset, the predictive power of the models mentioned above is reduced. As a result, a more sophisticated and robust computer model in bioinformatics is necessary to detect diabetic mellitus. Using an improved deep learning algorithm and fast feature extraction methods, this research provides a solid and trustworthy computational prediction for reliably recognizing diabetic mellitus. In this paper, we proposed a computational model using a multilayer DNN technique. The proposed model performance is thoroughly examined using the 10-fold cross-validation approach, which is based on several performance assessment indicators. The suggested predictor uses several modes of PseKNC as a feature formulation approach to construct the provided samples into a features set, including Single/Di/tri nucleotide composition. Second, an unsupervised PCA technique is used to eliminate unnecessary and noisy features, resulting in the selection of efficient features. The suggested predictor attained an average prediction accuracy of 88.65% based on the testing findings. Furthermore, the performance of the proposed predictor is compared to that of recently published predictors. The suggested model beat the current predictors in accuracy and other performance measures, according to the comparative findings. The major contributions of the paper are summarized as follow:

● Propose a deep generative model for classification of the probability of diabetes in patients at an early stage.

● The proposed model considered non-linearity in dataset using multi-stack processing layers and non-linear activation function.

● The performance of the proposed model is extensively evaluated using different performance measurement metrics such as Accuracy (ACC), Area under the Curve (AUC), Sensitivity (SN), Specificity (SP) and Mathew's Correlation Coefficient (MCC).

2.   The proposed model's design

The section outlines the suggested model's prototype and architecture. Figure 1 depicts the suggested model's design, which includes various elements that are addressed in-depth below.

2.1.   Benchmark dataset

A valid and trustworthy benchmark dataset is required to design a robust and resilient computational model. Therefore, we select a useful benchmark dataset available in ^{[16,17,18,19]} for training and validating the outcomes of the proposed model. Following Eq (1) represented the mathematical form of the chosen benchmark dataset.

The benchmark data were collected from medical and physical examinations and separated into healthy persons and diabetics. We used one of the physical examination data that contains 64,000 instances as the validation set. In the other dataset, 10,000 samples were randomly selected as an independent test set. We consider 14 different variables, i.e., age, pulse rate, breathe, left systolic pressure (LSP), proper systolic pressure (RSP), left diastolic pressure (LDP), right diastolic pressure (RDP), height, weight, physique index, fasting glucose, waistline, low-density lipoprotein (LDL), and high-density lipoprotein (HDL).

It is to be noted that the benchmark dataset is imbalanced (i.e., the positive samples are in the minority class, and the negative samples are in the majority class). The classification is biased toward the majority class in the imbalanced dataset, referred to as a classification problem, as shown in Figure 2. Various approaches, i.e., data-level approach and algorithm-level approach, have been considered in the literature for the class imbalance issue; however, the data-level system, including over-sampling and under-sampling methods, is widely employed the issue of the imbalanced dataset. The over-sample method could increase the risk of model over-fitting due to data sample replication. The under-sampling method can offer a modest solution in most cases. Hence, in this study, we applied the under-sampling method with the Near Miss procedure implemented in Python to balance the benchmark dataset.

2.2.   Feature formulation technique

The data is mainly analyzed in discrete form by quantitative machine learning techniques ^[20]. Formulating biological sequences into discrete forms is difficult in computational biology ^[21]. Various approaches ^[22,23] for transforming biological sequences into discrete feature vectors have been proposed in the literature. Chou's suggested approach has been widely adopted in current application development and is now used in all sectors of computational biology ^[24,25,26]. Chou's Pseudo K-tuple Nucleotide Composition (PseKNC) approach for representing RNA/DNA sequences into a discrete feature vector is widely used in computational biology ^[27,28]. We used the PseKNC approach to encode diabetes samples into a discrete feature set in this research.

To encode the following sequence D into a discrete feature vector, consider the following:

Here d1 represents the very first nucleotide in D, d2 represents the next nucleotide in D, and dL represents the last nucleotide in D. Equation (3) may be stated in the PseKNC's basic form utilizing Eq (4).

Where, T is the transpose, z is the integer number and ${\phi _u}$ is feature vector. The selected vector can be calculated by Eq (5).

Where, Eq (6) computed the coupling factor i.e., ${C_{i, i + j}}$

Finally, PseKNC technique to calculate the samples by using K = 1, 2, 3 in Eq (6).

2.3.   Feature selection

The feature vector may contain noisy, redundant, and irrelevant features significantly impacting a classifier's performance. Hence, in this paper, we perform feature selection using Principal Component Analysis (PCA) algorithm. The PCA algorithm ^{[29,30,31,32,33,34]} is a multivariate data processing method that computes covariance matrices and eigenvectors to reduce feature vector dimensionality to low-dimensional vectors. The PCA algorithm's primary goal is to decrease a feature vector's dimension while retaining as many discriminative features as possible. For instance consider i × j dimension as feature vector, "i" represent extracted features and "j" represent total samples. Using Eq (10) consider the input feature vector D:

Using the PCA algorithm, we took the following procedures to reduce the feature vector dimensionality.

Step 1: Using Eq (11), we first calculate the feature vector mean value

Step 2: Using Eq (12), from ${D_n}$ we subtract the mean value i.e., $\mathop D\limits^\_$

Step 3: Using Eq (13), covariance matrix C_n is computed

Where, ${B_{}} = {\left\{ {D_1^{'}, \mathop {}\limits^{} D_2^{'}, \mathop {}\limits^{}..., \mathop {}\limits^{} D_p^{'}} \right\}_{}}(i\mathop {}\limits^{} *\mathop {}\limits^{} j)$ and T = transpose of a matrix.

Step 4: Eigenvalues are computed in this step. In other words, the first Eigen must have a value larger than that of the second Eigen, and so on.

Step 5: Eigenvector is compute in this step as:

Last step: We choose k values with the highest Eigenvalues. Eigen vector with the highest Eigen values was chosen to reconstruct the new set of feature vector to accomplish a better prediction performance.

2.4.   Fully connected deep neural network

DNN is a branch of AI algorithms provoked by the functioning mechanism and activities of the human brain ^[35,36,37], as shown in Figure 3. Hidden layers are significant components of the DNN model and are actively involved in the learning process ^[38,39,40]. Using more hidden layers in the training phase can improve the performance of a model; however, it may cause significant issues such as model complexity, computation cost, and over-fitting ^[41,42,43]. The DNN model can extract relevant features automatically from the given unlabeled or unstructured dataset without the involvement of human engineering and expertise using standard learning procedures ^{[37,44,45,46]}. It has been reported by several investigators that the DNN model performed better than the conventional learning methods applied for various complex classification problems ^[47,48,49]. Moreover, deep learning algorithms are successfully applied in several domains, such as speech recognition ^[50,51], image recognition ^[52,53], bio-engineering ^[54,55], and natural language processing ^[56,57]. Inspired by the tremendous performance of deep learning in various domains for a complex classification problem, this paper has applied the DNN algorithm as a prediction engine.

Inspired by the effective employment of deep learning models in numerous domains for composite classification problems, this paper considered the DNN model to predict diabetes mellitus using a benchmark dataset. As shown in Figure 3, the proposed deep neural network method has six hidden layers, an input and output layer. Each layer in the DNN topology is configured with numerous nodes i.e., take input features set and produce output using the following equation.

Where, ${y_a}$ represent output at a layer $a$, $B$ represent bias value, $w_b^a$ mean weight used at a layer $a$ by a neuron $b$, $x_b^a$ represent input feature and $f$ represent a non-linear activation Tanh function and it can be calculated using Eq (16).

3.   Performance evaluation metrics

Any statistical machine learning-based model's performance can be evaluated using some measurement metrics before the model deploy in a natural production environment ^[58]. Several performance measurements have been proposed in the literature to assess a learning model. Accuracy is the primary criterion for evaluating the algorithm's effectiveness in every one of these measures. More than the accuracy metrics for a model performance evaluation is required. Hence, different performance metrics set is considered in the literature to evaluate a model ^[59,60,61]. These performance metrics, along with accuracy, are included: Area under the Curve (AUC), Sensitivity (SN), Specificity (SP), and Mathew's Correlation Coefficient (MCC). We considered all the performance as mentioned earlier measurement metrics to assess the proposed Deep-Sumo performance as these metrics were widely adopted in a series of publications ^{[25,46,62,63,64,65,66]}. The performance measurement metrics can be calculated using the following equations.

Where, ${D^ + }$ represent the total number of diabetes mellitus samples, ${D^ - }$ represent total non- diabetes mellitus samples. Similarly, $D_ - ^ +$ represents total mellitus that is wrongly recognized as non- diabetes mellitus, $D_ + ^ -$ represent total non- diabetes mellitus that wrongly recognized as diabetes mellitus.

4.   Experimental results

4.1.   Feature analysis

We first performed an analysis of the feature on the selected benchmark dataset. The success rates of Pseudo-DNC, Pseudo-TNC, and Pseudo-Tetra NC and fusion feature space on classification methods are listed in Table 1. Among all hypothesis learners, deep neural network using fusion features in the case of k-folds cross-validation test obtained the highest accuracy of 84.11%, specificity of 82.72%, the sensitivity of 89.11%, and MCC of 0.645. Similarly, using individual feature space of Pseudo-DNC accomplished the performance ratio of 82.03%, with an MCC of 0.641, a sensitivity of 85.66%, and a specificity of 78.89%, respectively.

The dimensions of the fusion features vector were lowered utilizing the feature selection approach to increase the performance of the suggested model even further (i.e., PCA). Consequently, the suggested deep neural network model's success rate increased dramatically, with an average accuracy of 88.65%, specificity of 89.41%, sensitivity of 81.50%, and MCC of 0.766. Integrating all selected traits we highlight, contributes to enhancing MCC value.

4.2.   Performance analysis

In addition, we examined the performance of the deep neural network model in terms of AUC. The AUC is a frequently used metric for evaluating the effectiveness of a classification algorithm, and it is calculated in the range of 0 to 1. In comparison to a predictor that computes a lower value, a classifier that computes a more significant rate is deemed the best predictor ^[66]. Figure 4 depicts the DNN's effectiveness in AUC utilizing various sequence formation strategies. Figure 4 shows that, compared to alternative sequence formulation approaches, the DNN method produced the most outstanding results when applying fusion characteristics. The proposed method, for example, generates the maximum AUC values utilizing fusion features 0.946 compared to the second maximum value of 0.891, using Pseudo-DNC. These findings revealed that the deep neural network model performs the most remarkable prediction by employing fusion features.

Furthermore, the suggested model highlights the issues of under fitting and overfitting, as well as how linear regression using polynomial features may approximate nonlinear functions. The graphic depicts the function we wish to approximate, a cosine function component, as shown in Figure 5.

As shown in Figure 5, a linear function (polynomial of degree 1) is insufficient to match the training data. This is known as under fitting. A polynomial of degree 4 nearly exactly approximates the underlying function. For higher degrees, however, the model will over fit the training data, i.e., it will learn the noise in the training data. The mean squared error (MSE) on the validation set is calculated; the more significant the MSE, the less probable the model generalizes successfully from the training data.

4.3.   Performance comparison with machine learning techniques

To authenticate the efficiency of our suggested framework, we accompanied comparative experimentations with several well-known methods. Table 2 shows the performance measure value for each categorization algorithm. Table 2 shows that the overall performance accuracy of all categorization techniques is greater than 70%. The DNN model, for example, had an average accuracy of 88.65% compared with other ML methods. Similarly, MCC, which shows the model stability, the DNN model had the highest score of 0.815. Moreover, Gradient Boosting and Extra Trees Classifier both show better accuracy for both testing methods, while the SVM-Linear Kernel model came in second last with a score of 72.49%. Furthermore, the Naive Bayes model performed the worst regarding gauging performance.

Based on the data, the primary reason of DNN model outperformed traditional learning approaches is the multi-stack processing. Whereas traditional learning methods employ single-layer processing, which is inadequate for commerce with a complex nature dataset with high nonlinearity.

4.4.   Performance comparison with the existing predictors

The performance of the proposed model was compared with the existing predictors, as mentioned in ^{[16,17,18,19]}. The mentioned latest methods build prediction models based on machine learning algorithms. The performance of our proposed model and the existing benchmark models are evaluated on benchmark datasets by using 10-fold cross-validation. For facilitating comparison, Table 3 shows the corresponding results obtained by the existing state of the art methods.

The results demonstrated that the proposed model outperformed the recently published predictors. For example, the proposed model achieved the highest accuracy, i.e., 88.65%, compared with the existing predictor achieving the second highest success rate, i.e., 87.26%. These results confirm that the proposed model performed superior to the existing models, with an average success rate improvement of 5.61%.

5.   Conclusions

The medical issues present during diagnosis determine how an individual with metabolic disease is classified, and people may not always neatly fall into a specific category. Some people, for instance, cannot differentiate between type 1 and 2 diabetes. Diabetes has a broad spectrum of clinical symptoms and disease progression. Considering the importance of PseKNC, PCA, and deep learning algorithms, this paper proposed a robust and accurate predictor to improve the performance of diabetes prediction. A hybrid feature vector with significant undertaking dimensional feature space was created by combining the feature vectors created using the various PseKNC techniques. PCA algorithm was applied to select only prominent features by eliminating noisy and irrelevant parts. Several machine learning algorithms have been applied for piRNA prediction; however, the DNN model produced the best performance using a 10-fold cross-validation test. In addition, the proposed optimized DNN classifier algorithm outperformed existing state-of-the-art models with an accuracy improvement of 5.61%. Furthermore, the proposed model might be utilized for a wide range of commonly used medications that raise insulin resistance and impair beta cell activity, potentially leading to diabetes mellitus in vulnerable persons.

Acknowledgments

The authors would like to thank the Deanship of Scientific Research at Umm Al-Qura University for supporting this work under grant code 22UQU4310396DSR60.

Conflict of interest

The authors declare that they have no competing interests.