A comprehensive review of machine learning techniques on diabetes detection

Diabetes mellitus has been an increasing concern owing to its high morbidity, and the average age of individual affected by of individual affected by this disease has now decreased to mid-twenties. Given the high prevalence, it is necessary to address with this problem effectively. Many researchers and doctors have now developed detection techniques based on artificial intelligence to better approach problems that are missed due to human errors. Data mining techniques with algorithms such as - density-based spatial clustering of applications with noise and ordering points to identify the cluster structure, the use of machine vision systems to learn data on facial images, gain better features for model training, and diagnosis via presentation of iridocyclitis for detection of the disease through iris patterns have been deployed by various practitioners. Machine learning classifiers such as support vector machines, logistic regression, and decision trees, have been comparative discussed various authors. Deep learning models such as artificial neural networks and recurrent neural networks have been considered, with primary focus on long short-term memory and convolutional neural network architectures in comparison with other machine learning models. Various parameters such as the root-mean-square error, mean absolute errors, area under curves, and graphs with varying criteria are commonly used. In this study, challenges pertaining to data inadequacy and model deployment are discussed. The future scope of such methods has also been discussed, and new methods are expected to enhance the performance of existing models, allowing them to attain greater insight into the conditions on which the prevalence of the disease depends.


Introduction
Given the growing population, it is necessary to develop systems to augment health and mitigate increasing concerns around the world. As scientific research continues to advance, the development of such system is becoming more efficient. Healthcare systems are designed to provide people with the requirements for good health and perform the detection and diagnosis of disease and conditions correctly with greater efficiency, as proposed in the conventional methods. In general patients are often highly concerned as to the quality of healthcare system and facilities available to provide treatment. The benefits of improvements in healthcare systems tend to affect people who have prevailing ailments more directly, and this group comprises the majority of the group of people affected by many diseases such as diabetes, blood sugar, and blood pressure issues [1]. According to the National Diabetes Statistics Report 2020, every 1 in 10 people in United States have diabetes, and new cases of diabetes 1 and 2 have significantly increased among young people. As health and healthcare form a critical pillars of a healthy society, it is necessary to use the capabilities of computational methods and artificial intelligence [2] to develop new methods for application in healthcare systems to promote a healthier society and reduce the risk of such diseases in our generations, further increasing the quality of life.
There has been a huge impact in the medical world with the advancement of technology. Health outcomes may depend on a matter of seconds for individuals who may not be able to reach a hospital or receive emergency treatment. Technology bridges this gap in distance and resources for all people to whom its benefits are extended. Various technologies have been developed using magnetic resonance imaging machines in video technology. Internet-based applications can provide patients with customizable services. After a few clinical visits, the remainder of the work can be fulfilled through high-tech services such as telehealth. Clinicians can communicate with patients through the Internet to better serve their needs [3]. One example of the use of video technology involves the provision of such mechanization in case of emergency to patients in trauma in rural and urban areas where clinical care may be unavailable [4]. Technology has the capacity to enable home healthcare with better productivity and security [5]. Data accuracy and availability have been proposed as among the most significant problems faced by hospitals, which involve maintaining and further processing patient data. Various algorithms in the field of machine learning and deep learning have been beneficially applied in practice to medical treatments. Some state-of-the-art ideas emerge from the massive implementation of technologies such as the creation of matching algorithms and natural language processing [6]. Data mining can be used to extract data directly, instead of relying on expert knowledge. These methods are considered to produce unique and distinctive patterns to create personalized plans for each hospital [7].
Diabetes mellitus (DM) is one of the most archetypal diseases worldwide. It is a disease that implies that a person's body systems are unable to work efficiently to use energy from food. There are four major types of diabetes known, including type-1, type-2, gestational, and other forms, with the most common being type-1 and type-2 [8]. Type-1 diabetes usually occurs in the young age group of 30-40 and is insulin dependent. Patients are required to take in doses of insulin for entire lives. In contrast type-2, predominates among the over-40 age cohort and is often related patients' weight. Type-2 diabetes is known to have greater prevalence globally, accounting for more than 90 % of cases [9]. Because this disease has long been near the top of global rankings listing serious of diseases, many researchers and doctors have proposed algorithms and methods for its treatment and detection. The implementation of these algorithms is rooted in the disciplines of deep learning and machine learning. With predictive analysis supported by neural networks, specifically convolutional neural networks (CNNs) and recurrent neural networks (RNNs), such methods have the ability to determine sentiments and learn high model features automatically [10,11]. Some researchers have implemented the use of machine learning through algorithms such as gradient boosted trees, which can be used to create predictive models of the progression from prediabetes to diabetes, and have aimed to provide early diagnosis for better treatment and to reduce further risks [12]. Another study applied a modified support vector machine (SVM) algorithm as an efficient method for both linear and non-linear data [13].
Since the advent of artificial intelligence and related technologies, such computational methods have been applied to real-time detection models in almost every field. The use of data mining, machine learning, deep learning, and computer vision has drastically reduced the difficulty of studying newer techniques that can significantly improvement of existing methods. In the next section, the algorithms and methods are surveyed.

Application of latest technology in diabetes detection
The algorithms used in data mining, machine learning, or any field of artificial intelligence perform predictive modeling, that is, the use of data and statistics to predict future outcomes based on historical data. The most common symptoms of diabetes include abnormal metabolism, hyperglycemia, and an associated risk for specific complications affecting the eyes, kidneys, and nervous system, which are major parts of the body. Such symptoms are used to gather data, and then the modeling is performed based on age and gender categories. One such algorithm is ordering points to identify cluster structure (OPTICS), which is set of ordering points to identify clustering structures. OPTICS is an advanced version of density based spatial clustering of applications (DBSCAN) with noise, and it eliminates all negative aspects of DBSCAN. The data clustering method used in this algorithm is a balanced iterative reducing and clustering algorithm using hierarchies (BIRCH) that selects the most suitable data for further analysis. Thus, the naïve Bayes (NB) data mining technique is used, and BIRCH and OPTICS are used for clustering similar types of data and used for identification of the correct algorithm for better accuracy [14]. Apache Spark is among the fastest growing platforms for health analysis. It operates more rapidly than the Hadoop platform, making it more easily usable and applicable to clinical practice [11].
Another such application in the field of OpenCV is the use of the computer-assisted non-invasive DM detection system. This system provided immediate results from facial images. The model examined four health blocks of the skin: forehead, left cheek, right cheek, and nose bridge. Then, feature extraction was performed using a local binary pattern and then classification using the k-nearest neighbor (KNN) clustering and a SVM. All these features were connected with software and show the results in real time [15].
Islam et al. [16] used different algorithms for a diabetes symptom dataset, include NB, random forest (RF), logistic regression (LR), and a decision tree (DT). First, the dataset with the patient system is entered into the system on which predictive algorithms are applied. After this, the dataset is input to the database and the performance or accuracy achieved by the model is observed. The most suitable algorithm is commonly selected based on the highest accuracy and best performance. The user's data is again taken as the input for the algorithm for further training and evaluation to increase the accuracy of the model in real time.
Irido-diagnosis is a predictive system in which the disease is detected through iris patterns. This was used for the detection of diabetes via the following method. Data were gathered from diabetic and non-diabetic patients on which pre-processing was performed. The next part is called image segmentation, wherein the iris of the eye is separated from the image in the dataset. Normalization is performed wherein the circular iris is converted to a rectangle using polar mapping. Feature extraction is then performed on this final image. A graylevel co-occurrence matrix is used to characterize images, which is a process for examining the texture. This process assigns numbers from 1 to 8 to images areas by calculating how often pairs of pixels with specific values and in a specified spatial relationship occur in the image. In this matrix, features such as contrast, correlation, dissimilarity, homogeneity, variance, and entropy are the features are pertinent to provide high-quality data.
These are just some of the authors among many who have contributed to the literature in this field. Below, a detailed survey of the methodology used by other researchers in this field is provided.

Machine learning in diabetes detection
Machine learning is a method by which a computational system learns the features of input data. Such methods haves proven effective for the detection of diabetes. Many machine learning algorithms have been developed, including supervised, unsupervised, and reinforcement learning methods. This is evidently practical because machine learning methods are driven by data. With such massive amounts of data fed into the database, machine learning can save considerable human effort. Models are trained on this data and provide the most suitable output based on the input data. The models can be trained on any parameters that are feasible for practicality and medical requirements. Some might examine facial features, while others look for blood report data obtained from patients. Because there are many symptoms of the disease, the parameters vary accordingly. With many proposed methods, researchers have probed various algorithms and tweaked numerous hyperparameters to obtain results that seem most suitable for real-life applications.
Choudhury and Gupta [17] used different algorithms to classify people into two categories: high-and low-risk individuals. They used a SVM to establish a hyperplane for categorization, a KNN classification technique for clustering new data into groups, DTs, RF and NB classifiers, and the binary classifier method called LR. On comparing the accuracies for this classification in the form of a confusion matrix, as shown in Fig. 1, the LR algorithm was found to be the most efficient and accurate, while the DT algorithm, achieved the lowest accuracy.
Shukla [18] used a LR algorithm, took out a dataset that showed the maximum accuracy would be yielded if parameters such as glucose, body mass index (BMI), and pregnancies, were used, which were represented in the form of a bar chart, as shown in Fig. 2. The author also attempted to showcase that the disease predominantly depends on those features that seem meager to us but have noted by doctors as relevant in possibly leading to a higher risk of diseases later. The LR model trained with the dominant features showed an accuracy of 82.92%. For the model forecasting, 0.458 was the probability of class zero and 0.572 for class one, which estimates the probability of a person being diabetic.
Dalakleidi et al. [19] used two datasets named PID, Case 1 and Hippokrateion, which is Case 2 from that the PID is split into 50% for training and 50% for testing, whereas the Hippokrateion has a bifurcation of 70% for training and the other 30% for testing. They used binary logistic regression (BLM), logistic model tree algorithm (LMT), which is a combination of LR and DT learning in simple models. The model's performance was measured using classification accuracy (ACC) and area under the curve (AUC). BLM achieved an ACC of 80.47 and AUC of 0.85, whereas the LMT achieved an ACC of   [20] used several algorithms to analyze a dataset using the NB and LR algorithms as well as the RF algorithm, after applying 10-fold crossvalidation and percentage split evaluation techniques. Figure 3 shows their proposed architecture. The dataset contained records of 520 people who were asked for possible reasons for diabetes. After data preprocessing, there were a total of 314 positive values and 186 negative values. Positive values represent the person being diabetic, and negative implies that they were not. The best result was achieved using the RF algorithm with an accuracy of 99%. Thus, it is an effective algorithm for a newly created dataset. Figure 4 shows exactly how each algorithm performed on modelling and prediction.
Harris et al. [21] performed clinical diagnosis for the detection of non-insulin dependent diabetes mellitus (NIDDM) using weighted linear regression. The relationship between the prevalence of retinopathy and duration of NIDDM was determined according to individual years of duration and assessed using weighted linear regression with weights for each year's data being inversely proportional to the binomial variance. The author stated that the retinopathy condition is an important parameter for the early diagnosis of the disease. It typically appears almost 4-7 years earlier than the clinical diagnosis of the disease. Figure 5 provides an accurate graph of the obtained results.
Ameena and Ashadevi [22] used the R language to build a model on SVM, DTs, RF, and LR. They used a dataset of 768 women, all of whom were older than 20 years. They used the following features: BMI, blood sugar, number of pregnancies, and diabetes pedigree function. They are defined two classes: 1, which affirmed diabetes and 0 for negation. On a comparison of the accuracies, the author concluded that the RF algorithm  showed the maximum correct estimations, with an accuracy of almost 77% compared to the other models. Daanouni et al. [23] used KNN and the DT algorithm on two datasets, with the first one having 2000 instances and the second having 768. They used eight features or attributes to train the model, such as BMI, glucose, blood sugar, and pregnancy. The authors used 80% for training and the remaining 20% for testing. They used optimized hyperparameters to reduce the loss. The results are plotted on two types of data: pre-processing and without. The comparison of results was performed using a receiver operating curve (ROC). The author concluded that KNN has a maximum accuracy of 97.53% and an AUC of 0.9689. Table 1 shows a comparison table for the accuracies obtained for training the model using the KNN classifier.
Sisodia D and Sisodia DS [24] used three classifiers, including SVM, NB, and DT. The classification is performed on PIMA Indian diabetes dataset, which is the PIMA Indian diabetes dataset taken from the UCI. To measure the accuracy, internal cross-validation was 10-folds. Accuracy, F-measure, recall, precision, and ROC measures were used. The attributes used were glucose concentration, blood pressure, BMI, age, skin fold thickness, number of pregnancies, 2-h insulin concentration, pedigree function, and class 0 or 1. On modeling, the authors computed that NB showed the maximum accuracy with 586 correctly identifying instances. The following Fig. 6 shows the different types of classifiers used along with number of classified instances.
Ahuja et al. [25] used the dataset from the UCI containing 768 records of women in which 500 were diabetic, while the remaining 268 were not. The authors used eight features for classification and applied a feature selection technique, which is linear discriminant analysis (LDA), to extract the important features required for classification. They used five types of classifiers for machine learning, including SVM, DT, LR, RF, and a multilayer perceptron. The authors used four parameters for evaluation, including accuracy, precision, recall, and F score. Based on these parameters, the authors concluded that multilayer perceptron yields the best results. Table 2 mentions the results using different values of k-fold validation.
Alehegn et al. [26] used the PIMA Indian diabetes dataset with eight features to train on and the 130 D hospital dataset with a larger number of values. There were four classification methods used, including RF, KNN, NB, and J48-DT algorithm. J48 is an upgraded version of the Iterative Dichotomiser 3 (ID3) classification algorithm. A 10 K cross-validation was used for 90% training and 10% testing. The author built a hybrid   model consisting of all of the above algorithms. The author concluded that NB and J48 are good for large data computations, and the KNN classifier is better for smaller datasets. Figure 7 shows the different algorithms used along with the correctly and incorrectly identified instances. Some more work done by other researchers which has been mentioned in Table 3. It contains a study of the machine learning algorithms used in the methods.

Deep learning in diabetes detection
Deep learning is a computational field that is usually involved where high computational power is required. Deep learning focuses on neural networks, their types, training epochs, layers of hidden, input, and output. The input layer is the first layer, and the hidden layers are responsible for all the calculations and manipulations, such as convolutions and pooling. The output layer determines the number of classes for the classification. Because of data augmentation, which means tweaking the data to increase accuracy, is also available in deep learning, it finds many applications with image training. The more layers the network has, the more it is capable of classification. Because of the many advantages, it has been widely used in the medical field to compute results with high accuracy. There are different types of networks with the most proficient artificial neural networks (ANNs), deep neural networks (DNNs), CNNs, and RNNs. Many researchers who work for the detection of a disease compare machine learning and deep learning algorithms to analyze which provides maximum accuracy.
Daanouni et al. [23] used ANNs and DNNs on two datasets of 2000 and 768 instances. They included eight attributes with the label of output as 1 for positive and 0 for negative results. The network was trained on two types of data: pre-processed and non-pre-processed. The DNN model seems to achieve high accuracy on both the data obtained, with an accuracy of 98% for the preprocessed dataset and 99.5% for dataset 1. On dataset 2, the non-pre-processed data had an accuracy of 80.99% and the other 96.35%. Hence, the authors concluded that DNN is an optimal classifier for diabetes detection.
Rakshit et al. [40] used R, SQL, and Python in a Microsoft Azure machine learning studio environment with the PIMA diabetes dataset, in which 80% was used for training and the other 20% for testing. This dataset is primarily concerned with diabetes in women. This contains eight attributes that are important for model building for a class -2 neural network. Figure 8 shows the general representation of the neural network. The hidden layer had 100 nodes, with the output layer connected to the nth hidden layer. With the model trained for over 1000 epochs with a learning rate of 0.01, they achieved an accuracy of 83.3% on a dataset with 262 negative cases and 131 positive cases.
Sapon et al. [41] presented diabetes prediction using supervised ANNs. The dataset comprises approximately 250 patients with 27 variables or features,, where exactly 50% was used for training, while the remaining 50% was used for testing with the MATLAB tool. The gradient algorithms used included the Fletcher-Powell conjugate gradient, Polak-Ribiére conjugate gradient, and scaled conjugate gradient. These algorithms were used to train the model and then analyzed using the correlation coefficient (CC) R. The results of these algorithms are plotted at different epochs against the mean square error. Based on a comparison of the values of R, the authors conclude that the scaled conjugate gradient confirms the highest accuracy with a value of 0.88, followed by the Fletcher-Powell conjugate gradient with a value of 0.097219 and, finally, the Polak-Ribiére conjugate gradient with 0.056466.
Refs. [31,42] both performed detection using an ANN. Ref. [42] used the PIMA Indian population dataset for women in Phoenix, while [31] performed this algorithm  The author claimed that these ensemble algorithms can be used on other disease datasets to increase accuracy.
The AdaBoost algorithm with the J48 as the base classifier showed the maximum accuracy followed by bagging and then the J48 classifier.
The AROC was used as the parameter. [27] NB with clustering Dataset used was the PIMA Indians Diabetes Dataset with eight attributes. The model is NB performed on prior clustering. This model is compared with only the NB model. Five hundred and thirty-one instances of data were divided into 5 clusters. The fourth cluster was the only one used for testing, which consisted of 148 instances.
By collecting a large amount of data for training, the accuracy can be increased by many-fold, helping people by developing a system that gives them a correct prediction without having to consult a doctor.
The parameters used for evaluation are accuracy, sensitivity, and specificity. The model with clustering showed a 10% increased accuracy, rise in sensitivity by 53.11% but the limitation caused here was the fall of specificity by 10.99% and also a reduced amount of dataset. [28] DTs, LR, and NB with bagging and boosting Initial datasets were collected from primary care units, which (through further changes) consisted of 11 features and a data of 30122 people. The three algorithms are used along with bagging and boosting methods, which are to decrease overfitting and increase accuracy.
The final model obtained with highest accuracy was deployed on a commercial web application. The author collected a raw dataset from Noakhali medical hospital containing 9843 samples with 14 attributes. Eighty percent of the data was taken for training and the rest for testing. All the algorithms chosen by the author were used for model building and then validation was performed on them using k-fold.
The author proposed that we can enhance the accuracy of early treatment to lessen the suffering of patients. Additionally, we can implement more classifiers to pick up the leading one for record-breaking performance and extend it to automation analysis.
The RF classifier was the algorithm that performed the best in classifying data and LR showed the worst performance. Although machine learning classifiers are widely used, they still lack in terms of accuracy against deep learning models. [30] LR and DTs The dataset was prepared using a questionnaire carried out for 1487 individuals in which 735 were diabetic and the remaining 752 negatives. A Pearson chi-square test was carried out on all the characteristics. The models' performance was evaluated on three parameters: accuracy, sensitivity, and specificity. Apart from this, a confusion matrix was also built to determine model performance.
Recently, many researchers have been implementing various algorithms and networks to compare them and find out the most feasible one. DTs and LR are among the ones that are most used. On using a smaller subset of features, the LR model performed slightly better than the SVM model. [32] DTs and NB The dataset taken for consideration was the PIMA Indian diabetes database. On applying feature selection, the author obtained five features. 10-fold crossvalidation was used for data preparation after which the J48 algorithm -DTs and The author proposed to gather information for the dataset from different people to make a more representative model. The work can be further enhanced to include automation.  References NB is applied. The model performance was evaluated using mean absolute error (MAE), root-mean-square error (RMSE), relative absolute error, root relative squared error, and kappa statistic.
the models are not showing good accuracy on larger datasets.

RF and XG boost
The author used the PIMA diabetes dataset. Using Jupyter Notebook as an IDE, the author trains the model using 8 attributes of the total 9 provided in the dataset. The algorithms used were RF and XGBoost. After setting the hyperparameters the models were trained.
The author suggests the use of more algorithms in this branch of machine learning like hybrid model for better accuracies.
The accuracy gained on the RF classifier came out to be 71.9%. The hybrid model proposed through XG boost gained an accuracy of 74.1%. The accuracies gained on the models were comparatively less compared those already available. The hyperparameter tuning needs to be set better for optimizing the algorithm. [34] DT (J48) and NB The author used the PIMA Indian diabetes dataset with 8 attributes, which was reduced to 5 based on the feature selection. The pre-processing was performed used the WEKA using 10-fold validation. The model was created using the 70% dataset and the rest was used for testing.
In future, it is planned to gather the information from different locales over the world and make a more precise and general prescient model for diabetes conclusion. Future study will likewise focus on gathering information from a later time period and discover new potential prognostic elements to be incorporated. The work can be extended and improved for the automation of diabetes analysis.
The J48 algorithm was 76.95% accurate with other parameters like kappa statistic, MAE, RMSE, relative absolute error, and root relative absolute error. The NB algorithm was accurate up to 79.56%. Since this model is not optimally configured, a developed model would require more training data for creation and testing. [33]

Genetic algorithms with fuzzy logic
This work is a model implementing the genetic and fuzzy algorithms for effective disease prediction. For the implementation of GA, MATLAB R2006b was used. The principle of feature selection was implemented using fuzzy logic algorithms. Firstly, a simultaneous mapping was performed based on an appropriateness measure of variables values to each class using suitable membership functions according to each type of feature. Then, simple fuzzy reasoning mechanisms were proposed to deal, in unified way, with classification.
The proposed work helps minimize the cost and increase accuracy and can be used in future for better implementations.
Through this approach of GA, the accuracy went up to 87% with the training cost reducing by more than 50%. [35] k-NN, NB, DT, RF, SVM LR These models were created in comparison for detection of type-2 diabetes. A total of 300 samples were taken in which 161 were diabetic, 60 non-diabetic and the rest were unconfirmed. Using feature summarization, eight features were selected with the WEKA tool used. These algorithms were used against a proposed framework that automatically extracts patterns of type 2 DM.
An application of genome wide association and phenome-wide association study in hope for its associations with DM. They proved to be an important association for future models.
Proposed model was evaluated on basis of accuracy, precision, specificity, sensitivity, and AUC. The proposed algorithm gained an AUC of 0.98, which outperforms the stateof-the-art AUC of 0.71. [36] LR, DT, RF, SVM The author used the PIMA Indian women dataset concerned with women's health with 8 attributes. Different models were trained for this dataset under different hyperparameters.
The author proposed to create advanced models on RF because of its highest accuracy and ability to overcome overfitting.
Different models were compared on basis of accuracy. RF gained the highest accuracy with 77.06% followed by SVM. [22] DT -J48, RF The author obtained dataset from Luzhou from hospital physical examination. An independent test set was taken with 13700 samples. The data contained 14 attributes. Another dataset was the PIMA Indian diabetes dataset. The DT and RF algorithm were implemented in WEKA with principal component analysis.
The author hoped to predict the type of diabetes using a dataset containing the required data which would lead to be an added advantage for improving the accuracy. Ref. [43] conducted a comparative study of neural networks in diabetes detection. Using the PIMA Indian diabetes set again, they used the eight common features required for model preparation. With the first 576 cases used for testing, they used a 10-fold cross-validation technique too estimate the results. The author used a multilayer neural network (MLNN) with 50 neurons for each hidden layer and an output layer using the nonsigmoid activation function and a PNN, which is a probabilistic neural network with a single hidden layer. The author concluded his results on accuracy, showing that the MLNN model achieved an accuracy of 79.62, and the PNN achieved an accuracy of 78.65.
Ref. [44] used the PIMA diabetes set and computed the model using an ANN and used eight features for modeling. The author explained the different functions used for pre-processing and model training. The activation function was a sigmoid function, and backpropagation is used to calculate the gradient of the loss function. The error function computed the final error to be approximately 8% at the end of the model building. The results were validated on ROC and RMSE. The author achieved an ROC area of 0.88 and RMSE equal to 0.39, which is a FAIR classifier. Figure 9 shows the results as plotted in form of lines.
Ref. [29] used an ANN for his model trained over a dataset consisting of over 30000 instances and 11 features to train on. With the hidden layers equal to 12, the values of the layers were calculated using the sigmoid function. Bagging and boosting methods were implemented. Bagging was also set to reduce the variance in the model, and boosting was performed to reduce the error. The neural network with bagging achieved an accuracy of 85.324%, followed by the ANN model with boosting with an accuracy of 84.815%, and then the ANN with 84.532%. Finally, the author validates the final comparison using an ROC.
Ref. [45] used machine learning and deep learning techniques to detect DM. The dataset used here was the PIMA Indian diabetes dataset consisting of 768 features and eight features to train on with a total of 500 instances belonging to the non-diabetic class and the remaining 268, which are diabetic class. Sixty percent of the data was selected for training and the remaining 40% for testing; a CNN was used. They are composed of three layers, with the classification is performed by the output layer. The prediction accuracy achieved using the model was 76.81%. Some more work done by other researchers which has been mentioned in Table 4. It contains a study of the deep learning algorithms used in the methods. Classifiers of this kind can help in early detection of the vulnerability of a diabetic patient to heart disease. There by the patients can be forewarned to change their lifestyle. This will result in preventing diabetic patients from being affected by heart disease, thereby resulting in low mortality rates as well as reduced cost on health for the state.
NB was able to classify 74% of instances correctly. For SVM, the accuracy gained was 95.6%. [38] DT, SVM, K-NN RF The author used the PIMA Indian dataset. Then, data normalization was done followed by feature selection.
The benefit of this optimized machine learning model is that it is suitable for patients with DM and it can be applied to all the fields of health care environments.
The models are compared on the basis of accuracy, sensitivity, and specificity. DT has 78.25% accuracy. [39]

Challenges and future scope
Although there many methods and algorithms which have been proposed in the field of machine learning or deep learning, many challenges remain, as mentioned by the authors in their works. The first concern that comes to mind when building a model is data. Refs. [32,33,49] were some of the authors that referenced the problem. One of the biggest problems encountered during the survey of the papers was finding articles and papers that did not relate to the popular PIMA Indian dataset. This dataset was chosen to ensure that the models provided good results because of the establishment of the dataset. The datasets are either too small or inadequate, or they lack real-time data. Small datasets pose a problem of overfitting on the model, which shows higher accuracy, but they are not able to deal with newer testing data. Hence, the model is not feasible for real-time implementation. Some authors dealt with CGM data, which was real-time, but the model training with that data was not efficient. The datasets were also selected from particular regions that are not representative of a common system. Different regions have different people and lifestyles. Generally, researchers spend 80% of their time cleaning and managing data for model training. Hence, data complexity leads to higher cost and maintenance charges. The next step is feature selection. While some authors neglected some of the features, some grouped them for feasible training. Every dataset poses the problem of having appropriate features to cater to the needs of a single algorithm. After all the data are made available, the technical stacks are finalized. Many tools are available to construct machine learning models, but choosing a model to optimize performance is also necessary. The next challenge is debugging. This becomes easy if tools such as Jupyter Notebooks are used where the code is divided into cells. This becomes difficult when the model runs on automation batch processes. In addition, as there are only a few diabetes datasets available on the Internet; more public data should be available for research. More research should be performed using heart rate, as it requires less bandwidth, and its computational complexity is also low. They can also be used in cloud or mobile devices. HR signals should also be used to detect other cardiac diseases. In some cases, authors required a time-series dataset. Since they are not available across any online resource, it is difficult to replicate such work. Such special models require extensive tuning and large datasets for both training and testing. The next part is the construction of an actual model. To achieve perfect accuracy, many parameters must be adjusted. Random states, kernel, number of trees, hyperparameter tuning, and various others are considered while creating a model. Selecting a correct algorithm with suitable hyperparameters should also be performed precisely. Some classification models will only train on a single parameter, which results in a decreased accuracy for the model in real-time detection. It is evident from the analysis of these schemes in all classes that most of them suffer from either a single data input parameter or the feature selection is not optimal. Along with such restrictions on parameters, few classification-based schemes are purely dependent on kinds of hardware devices, which increases the difficulty of availability and adaptability of these schemes.
A healthcare-based machine learning model is only useful if it can be used for the benefit of people. Here, the model deployment in practical applications is critical. Many authors have proposed deploying models on mobile platforms. In real-life implementations, only engineers with background and experience with cloud servers and DevOps can deploy models. In this ongoing process, many issues need to be considered, such as how frequently the predictions are required to be displayed or the number of applications that are required for model processing. Although considerable precautions were taken to ensure there were no discrepancies in the study, no study could claim to be perfect and there is always scope for improvement. The development of more inquisitive study providing deeper insights into aspects that enable the predictive power of models rather than only pre-defined parameters such as accuracy, precision, F1 score, ROC, and AUC would be beneficial in the future. For classification, to distinguish between diabetic and normal profiles, clustering-based schemes provide accurate results. However, most of the clustering algorithms struggle with plug-n-play problems, which means that they usually contain human intervention during classification and analysis, which involves the possibility of error.

References
Deep belief neural network The dataset used was PIMA Indian diabetes dataset with 768 instances and 8 features. The activation hidden function used was ReLU rectifier linear unit with three hidden layers and sigmoid as the input activation function. The batch size at 100 and epochs set to 5.
As the author compares his network with conventional methods of machine learning classifiers and it obtained high results, they believe that this network can be tweaked according to convenience and used for detection of other diseases as well.
The results of the network proposed by the author and conventional methods were seen on three parameters: recall value, precision and F1 measure. This network obtained high recall, precision, and F1 measure values, which shows that it is a good model.  [52] AlexNet and GoogLeNet In this study, the author developed a model called IGRNet to detect and diagnose prediabetes effectively using a 12-lead electrocardiogram lasting 5 s of 2251 case data. The neural networks used were compared with traditional ML algorithms like SVM, RF, and KNN.
The author suggested to use this hybrid model for future predictions due to its efficient performance.
The results of the networks were compared over different activation functions. The IGRNet model gained an accuracy of 85.6%, which was higher than any other model used for comparison. [53]

DNN
The data collection was retrieved from the UCI machine learning repository -PIMA Indian diabetes dataset. For the neural network, the hidden layer count is 4. The numbers of neurons in the layers were 12, 16, 16, and 14.
The proposed system will be supportive for the medical staff and as well as for the common people because with five-fold crossvalidation, the accuracy was more than any other model and on comparing it with other established models of authors, it came out to be the highest.
The data samples were divided into 5-fold and 10-fold cross-validation. The accuracy gained on this model for five-fold was 98.04% and for tenfold was 97.27%. The results were also analyzed through ROCs and F1 score.  With the ReLU activation function in the input layer and hidden layers, the output layer is sigmoid.
The author proposed to extend this work for further accuracy increase to help in early prediction for diabetes.
The results for the perceptron were calculated for 10 runs for 150 epochs. The highest average accuracy was gained by the network having nodes 8-128-128-1 followed by 8-64-64-1 and then 8-32-32-1. [56] ELM algorithm The data was acquired from multiple sources, including medical laboratories, hospitals, and public datasets. The dataset was preprocessed and meant to include 12 The goal of this research was the study of diabetic treatment in healthcare using big data and machine learning. It presents a big data processing system by employing The proposed approach was proven to be efficient. The goal was to reduce the FP and FN and boost the precision and recall rates, which was achieved by the author.
[57] Considering all the above challenges, it can still be considered that they can be overcome in the future. Scholars and clinicians will continue to work toward the construction of larger and better datasets and design more efficient models and algorithms for better classification and accuracy. Any of the diseases occurring on a wide scale, such as diabetes, can be controlled through artificial intelligence techniques and automation. One can create state-of-theart efficient models based on studies that provide early detection of diabetes and can help people to further change their lifestyle. Because deep learning performs better on most datasets, it should be combined with different algorithms to achieve better accuracy and performance. Hybrid schemes play an important role in improving the performance of the models. Through early detection, patients can be treated much earlier to avoid further risks of heart problems in cases of diabetes. Any model that can be deployed on mobile platforms should cater to the masses for their help and be representative. An implication of this survey is that ML models that have yielded efficient results that can be utilized by future researchers to further polish and improve as well as create a pipeline or an ensemble of correct and efficient models to increase the chances of predicting the disease with even more probability. Such models can be further improvised to automate the system created so that it can deal with newer data without problems.

Conclusions
Diabetes can be devastating after a certain period if not detected or diagnosed correctly. Many machine learning methods have been discussed, starting from different basic algorithms such as the LR, SVM, DTs, to further classification including the ID3, C4.5, C5.0, J48 and CART and NB. Ensemble methods, such as bagging, boosting, and RF regressors, are further used to enhance the accuracy and performance of models. These techniques have been implemented on all types of platforms such as Python or MATLAB, and the models have been analyzed using different parameters such as area under curve or confusion matrices or mathematical terms such as the RMSE or MAE. Machine learning has been introduced in medical diagnosis systems as it has proven to be accurate in detection, successful in application to treatments, and is more cost effective. Although the above are very strong classifiers, we believe that deep learning, which is a subset of machine learning, can learn large amounts of unstructured and unlabeled data. Deep learning models are more complex and accurate. Different models for deep learning start from the most basic ANNs to convolutional nets to further RNNs, including LSTM and Bi-LSTM. Temporal and deep belief networks have also been discussed. In contrast, deep learning involves some shortcomings such as increased computational time, resources, and frequent adjustment of the parameters. Deep learning performs better on image datasets; therefore, for diabetes diagnosis, images would be better. Most researchers have implemented several algorithms in both machine and deep learning to compare their performance on the data, while others have combined two or three methods to gain more accuracy on a single system. an ELM algorithm.
Neural network The author obtained dataset from Luzhou from hospital physical examination. An independent test set was taken out with 13700 samples. The data contained 14 attributes. Another dataset was the PIMA Indian diabetes dataset. This method is a two-layer network with sigmoid hidden and SoftMax output neurons.
The author hopes to predict the type of diabetes using a dataset which is a lead for improving accuracy.
The neural network gained an accuracy of 74.14% on Luzhou dataset and 74.75% on PIMA Indian dataset with the use of principle component analysis. [37] CNN Dataset used by the author was taken from National Institute of Diabetes, which consists of nine parameters. Fuzzification was done on the dataset for CNN which becomes populated.
The α values of the neural network were 2 and 5. These networks were compared with a CNN.
Fuzzification proves to be of a useful nature since it provides diverse data to train on.
The neural network with value of α = 2 performed better α value 5. CNN performed better than both of them. [58] Sharma and Shah Visual Computing for Industry, Biomedicine, and Art (2021) 4:30 Page 13 of 16