Machine learning (ML) is a type of artificial intelligence (AI). ML uses historical data as input to predict new output values [3] . Machine Learning has four main categories:

● Supervised learning: method that feeds input data as well as correct output data to the ML model. The aim of a supervised learning algorithm is to find a mapping function to map the input variable (x) with the output variable (y).

● Unsupervised Learning: method in which models are not supervised using a training data set. Instead, the models find hidden patterns and ideas on their own from the given data. It can be compared to the learning that takes place in the human brain when it learns new thing.

● Semi-supervised Learning: is the middle ground between supervised and unsupervised learning and operates on the data that consists of a few labels. It mostly consists of unlabeled data.

● Reinforcement Learning: feedback-based Machine Learning technique in which an agent learns to behave in an environment by performing the actions and seeing the results of actions. For each good action, the agent gets positive feedback, and for each bad action, the agent gets negative feedback or penalty.

Deep Learning is a component of Machine Learning methods based on Artificial Neural Networks with representation learning. The adjective “deep” in Deep Learning refers to the use of multiple layers in the network as shown in Figure 1.

There are many kinds of Deep Learning architectures such as Deep Neural Networks, Deep Belief Networks, Deep Reinforcement Learning, Recurrent Neural Networks, Convolutional Neural Networks and Transformers.

A Convolutional Neural Network (ConvNet/CNN) is a Deep Learning algorithm which can take in an input image, assign importance (learnable weights and biases) to various aspects/objects in the image and be able to differentiate one from the other [4] . The pre-processing required in a ConvNet is much lower as compared to other classification algorithms. While in primitive methods, filters are hand-engineered, with enough training, ConvNets have the ability to learn these filters/characteristics. Figure 2 describes how Convolutional Neural Networks work.

Cross-validation randomly partitions the training data into folds. The algorithm uses 10 samples by default if the dataset has not already partitioned. To split the dataset into a different number of samples, you can use the partition and sample component and specify the number of samples to use. Cross-validation is an essential tool in the Data Scientist toolbox. It allows to better use available data. When building a Machine Learning or Deep Learning model using some data, data are often divided into training and validation/test sets. The training set is used to train the model, and the validation/test set is used to validate it on data it has never seen before. The classic approach is to do a simple 80% - 20% split, sometimes with different values like 70% - 30% or 90% - 10%. In cross-validation, researchers can do more than one split. It can be 3, 5, 10 or any k number of splits. Those splits are called folds, and there are many interesting strategies to create these folds [5] . An example is given in Figure 3.

Deep Neural Network (DNN), an unsupervised learning approach, is used in [6] for a prediction on the Pima Indian diabetes dataset. The model achieved 98.16% accuracy. Deep learning can outperform shallow networks [7] .

Authors in [8] proposed a strategy for the diagnosis of diabetes that uses a Deep Neural Network by training its attributes by cross-validation at five and ten times. The accuracy obtained is 98.35%.

Moreover, authors in [9] proposed a model that uses the binary cross-entropy loss function in addition to a number of parameters. The prediction model that uses Deep Neural Network on the PIMA database gives performance that is about 99.41%.

In [10] , the criteria of authors were to minimize the error function in the training of the neural network using a Neural Network model. The test accuracy is 87.3%.

Besides, researchers in [11] collected electrocardiograms (ECGs) from 20 people with diabetes and a group of people lying in a stable posture for 10 minutes. Heart rate data are extracted from the ECG signals for the detection of diabetes, so the authors used 5-layer CNN, LSTM and 5-fold cross-validation and obtained 95.7% accuracy.

Authors in [12] used VRC (Heart Rate Variability) as data and CNN-LSTM (LSTM = Long Short Term Memory) to automatically detect the anomaly. 5-fold cross-validation and CNN are used and gave an accuracy of 93.6% while the CNN-LSTM combination gave the maximum accuracy of 95.1%.

Moreover, an approach is proposed [13] , based on convolutional long-term memory (CLSTM), diabetes classification model was developed and compared i with existing methods on the Pima Indians Diabetes Database (PIDD). The result obtained by the CLSTM model is higher than the other methodologies 96.8%.

In addition, authors in [14] present a methodology for diabetes prediction using various Machine Learning algorithms using the PIMA. DL and DT dataset provide promising accuracy (98.07%).

Finally, authors in [15] presented DeepCare, an end-to-end dynamic Deep Neural Network that reads medical records, stores the history of previous illnesses, infers the current disease state and predicts future medical outcomes. The performance obtained is 79%.

Early diagnosis of diabetes has been a topic in which Machine Learning algorithms have played a large role before the rise of Deep Learning.

Machine learning (ML) algorithms and Neural Network (NN) methods are used in [16] . Authors found that the model with logistic regression (LR) and Support Vector Machine (SVM) works well for diabetes prediction with an accuracy of 88.6%. Six learning-based classification methods were implemented and tested on a dataset collected through online and offline questionnaires with 18 diabetes-related questions. The same algorithms were also applied to the PIMA database. The experimental results show that the accuracy of Random Forest for our dataset pima is 94.10%, which is the highest among others.

Many Machine Learning algorithms are trained on many datasets in [17] . It appears that SVM outperforms the other algorithms with 98.6% accuracy.

On other hand, researchers in [18] proposed a new approach to efficiently predict diabetes from the medical records of Indian Pima patients. The modified J48 classifier was used to increase the accuracy rate of the data mining procedure. The WEKA data mining tool was used as an API of MATLAB to generate the modified J-48 classifiers. The experimental results showed a significant improvement over the existing J-48 algorithm. It was proved that the proposed algorithm can achieve 99.87% accuracy.

The soft voting ensemble classifier proposed in [19] uses the set of three Machine Learning algorithms, namely random forest, logistic regression and Naïve Bayes, for classification, taking accuracy as evaluation criteria. The proposed ensemble approach gives the highest accuracy with 79.04% on the PIMA diabetes dataset.

In addition, the objective of authors [20] is to establish the best Machine Learning classifier for the application of wafer defect detection. The experiments proved that the logistic regression classifier is the best classifier for detection with an accuracy of 86.9%.

Three Machine Learning classification algorithms, namely Decision Tree, SVM and Naive Bayes, are used in [21] . The experiments are performed on the Pima Indians Diabetes Database (PIDD). The results obtained, show that Naive Bayes outperforms the other algorithms with the highest accuracy of 76.30%. These results are verified using ROC (Receiver Operating Characteristic) curves.

Besides, authors in [22] provide a comparative analysis of different machine learning models to reach the most supporting decision for diagnosing heart disease with better accuracy as compared to existing models. The comparative study has proven that the XGB is the most suitable model due to its superior prediction capability to other models with an accuracy of 91.6% and 100% on two different heart ailments datasets, respectively.

Moreover, authors in [23] provide the comparative analysis of different Machine Learning algorithms for diagnosis of different diseases. It brings attention towards the suite of Machine Learning algorithms and tools that are used for the analysis of diseases and decision-making process accordingly.

Futhermore, authors in [24] applied and compared Machine Learning techniques initially to predict the outcome of TB therapy. After feature analysis, six algorithms including decision tree (DT), Artificial Neural Network (ANN), Logistic Regression (LR), Radial Basis Function (RBF), Bayesian Networks (BN), and Support Vector Machine (SVM) are developed and validated.

Finally, in [25] significant attributes selection was done via the principal component analysis method. Their findings indicate a strong association of diabetes with body mass index (BMI) and with glucose level, which was extracted via the Apriori method. Artificial Neural Network (ANN), Random Forest (RF) and K-means clustering techniques were implemented for the prediction of diabetes. The ANN technique provided a best accuracy of 75.7%, and may be useful to assist medical professionals with treatment decisions.

The unrelated information will be removed by applying preprocessing. Pima

Indian Dataset (PID) consists of noise and empty entries, therefore preprocessing steps is used to reduce noise by eliminating redundant, empty, or any ambiguous data in the dataset. The dataset consists of 768 occurrences with nine attributes of features. It supports classification in a binary format that consists of 0 and 1 for diabetes positive and diabetes negative respectively. Diabetes positive is to identify persons with diabetes and diabetes negative is to identify persons without diabetes. In 768 in-stances, there are many entries with 0.

Multicollinearity is a condition where a predictor variable correlates with another predictor. Although multicollinearity doesn’t affect the model’s performance, it will affect the interpretability. If multicollinearity is not removed, it will be impossible know how much a variable contributes to the result. Figure 5 shows the correlation matrix of our model.

Feature selection phase improves the Deep Learning process and increases the predictive power of algorithms by selecting the most important variables and eliminating redundant and irrelevant features.

When a model is trained using all of the data in a single, it could not give the best performance accuracy. To resist, this k-fold cross-validation that helps us to

build the model, is a generalized one. To achieve this k-Fold cross-validation, the dataset is divided into two sets: training and testing, with the challenge of the volume of the data. Here Test and Train dataset supported building model and hyperparameter assessments. The model has been validated multiple times based on the value assigned as a parameter and which is called k and it should be an INTEGER.

Based on the k value, the dataset would be divided, and train/testing will be conducted in a sequence way equal to k time. In this work k = 10.

Pima Indian Diabetes dataset consists of 768 tuples of instances with nine feature attributes. The diabetes dataset is suitable for binary classification models, where it has only binary values like 0 and 1, i.e., 0 for diabetes negative and 1 for diabetes positive. Out of nine (9) feature attributes from the dataset, only four (4) are considered for classification based on the feature Importance attribute score. The description of the PID dataset is shown in Figure 6.

The early prediction of diabetes is crucial and obtaining higher accuracy rate in diabetes prediction is decisive. All experiments are run on GPU enabled TensorFlow with Keras. Figure 7 shows the feature importance of this model.

In experimental analysis, the pima indian dataset have been divided between 80% - 20% for training and testing purpose. We obtained 89% of accuracy on 10 k-fold cross-validation. The results of comparative analysis showed that DNN could be able to give best accuracy for diabetes prediction on pima indian dataset but using k-fold cross-validation on DNN model and pima indian dataset could decrease the accuracy of model.

Through the confusion matrix, presented in Figure 8, the performance of the DNN classifier is validated. Accuracy is the prior most metric used to know the performance and effectiveness of the model.

The formula for calculating accuracy is (TP + TN)/(TP + FP + FN + TN) or all true positive and true negative cases divided by the number of all cases.

TN—True Negative, TP—True Positive, FP—False Positive, FN—False Negative.

● True Negative—Predicted negative, if person not affected by diabetes.

● True Positive—Predicted positive, if person affected by diabetes.

● False Positive—Predicted positive even if person not affected by diabetes.

● False Negative—Predicted negative even if person affected by diabetes.

Our main objective is to compare DNN models which are built on Pima Indian Dataset and the model that we proposed in this work and prove the impact of 10 k-fold cross-validation. Figure 9 and Figure 10 present the results of comparative study of five models on diabetes prediction using DNN on pima dataset. Table 1 gives precisions on performance metrics values for each model.

On Figure 9, the highest accuracy is seen on DNN + DT [14] but the highest sensitivity is in DNN [8] . However, specifivity and sensitivity of our proposed model, are less important on Figure 9. On other hand, Figure 10 shows that DNN on PIMA dataset [9] is the most performant model with the highest accuracy between the five models. Our proposed model tested with k-fold cross-validation is the less efficient.