Jorjani Biomedicine Journal

Introduction
According to the world health organization, diabetes is a chronic disease. This disease is due to either the pancreas not producing enough insulin or the body not responding properly to the insulin produced.
Diabetes has a serious global economic impact. In 2012, this disease with 1.5 million deaths, made the 8th leading cause of death. Based on statistics of 2017, nearly 425 million people had diabetes, about 2-5 million people die every year from diabetes disease. By 2045, it will reach 629 million (1).
This epidemic disease tends to grow. It was estimated that approximately 642 million people will have diabetes by 2040. However, the number of studies estimated that the annual cost of the disease from US$673 billion in 2015 reaches US$827 billion in 2016, which currently exceed the 2030 prediction (2).
Moreover, the economic burden of diabetes in China has risen over time. In 2015, the total annual expenditure was between US$51.1 and US$88.4 billion (3).
Type 1 diabetes is also called “insulin-dependent diabetes”. Type 1 diabetes results from the body’s inability to produce insulin. Therefore, the injection of insulin is essential for the patient. Type 2 diabetes occurs when the body’s cells do not react to insulin properly. This type of diabetes appears most often in middle-aged, and sometimes older people. Gestational diabetes is high blood sugar that develops during pregnancy in a woman who has not been diagnosed with diabetes before.
Today, the disease is also occurring in children. There are several causes of diabetes, including allergy, genetics, body weight, diet, physical activity, and, lifestyle (4).
Since diabetes is a multifactorial disease, it is diagnosis is complicated. Therefore, for accurate diagnosis, the physician should examine the test results of a patient, compare the test results with those of patients under the same condition, and analyze previous decisions. To care for a person, doctors are well prepared to identify people at risk for type 2 diabetes.
However, as we try to screen thousands of patients at high risk, the challenges that physicians face become apparent (26). and Analysis of the factors influencing the diagnosis may be influenced by human error. Because it is subject to the physician’s interpretation.
A Blood test is not discriminatory enough, and it is not enough for an accurate diagnosis of diabetes. Because among different characteristics of people it may be interpreted differently (4). Therefore, apart from laboratory information, lifestyle, physical condition, and family history of patients should be examined.
Despite human error probability, different types of machine learning algorithms, analytical and statistical methods, that use the previous and current data for finding information and predicting future events, can be used for solving this problem.
The artificial intelligence field is growing rapidly. Its functions in the field of diabetes are diagnosis and management of this chronic disease. The techniques of machine learning are used for making models of diabetes diagnosis and its side effects (5).
Results of empirical research articles, reflect the ability of analytical techniques to diagnose diabetes. Recent research shows that machine learning techniques can describe the characteristics of patients and identify patients at risk for type 2 diabetes (27,28). Machine Learning is one of the most important features of artificial intelligence that supports the development of computer systems with the ability to learn and improve from experiences, and the need to plan for every case (6).
The purpose of this study is to use learning machine algorithms for making models of type 2 diabetes diagnosis.
Related work
More than half of all adults worldwide are suffering from diabetes disease. Hence, early detection of the disease greatly improves patient quality of life. For the diagnosis of diabetes, a lot of researchers used learning machine methods and data analysis (7).
Many articles by using statistical analysis methods explored connections and samples between Chinese patients’ data (8-11).
Gao et al. in (8) used statistical analysis and multivariate Cox Regression algorithm to study the relationship between ALT and diabetes. They concluded that people between the age group 30-40 with elevated ALT are at a greater risk compared to those with low ALT.
Chen et al. in [9] by using statistical analysis and multivariate Cox Regression algorithm demonstrated that ratio TG /HDL-C  was positively associated with the incidence of diabetes in the Chinese population. 
Qin et al. in [3] by exploring ratio TG/HDL-C as an independent predictor of diabetes, found that TG/HDL-C is a powerful predictor in male patients compared to female patients.
Chen et al. in (10) investigated the relation of obesity and age with diabetes. Eventually, across all age groups, there was a linear association between BMI and diabetes. They verified that BMI and age are a strong predictor of diabetes.
Lin et al. in (11) designed a Nomogram based on the seven factors of diabetes in order to predict type 2 diabetes risk among people.
Diabetes prediction should be under control. Therefore, supervised learning algorithms have attracted a great deal of attention from researchers. By using machine learning, many diabetes research articles have been done on the Pima data set (12-15).
Wu et al. in (12) applied the Nearest Neighbor and Regression Logistic algorithms to make a diabetes prediction model with an accuracy of 95.42%.
Viloria et al. in (5) developed a diabetes predictive model by applying the Support Vector Machine algorithm. The algorithm was obtained with an accuracy of 99.2% with Colombian patients. 
Devi et al. in (17) by employing a Support Vector Machine algorithm, managed to make a diabetes prediction model with an accuracy of 99.4%.
Many researchers also by analyzing different types of machine learning algorithms, assess and compare their performance on a particular data set, and, ultimately with particular evaluation methodologies, introduce the most appropriate predictive model for diabetes (18, 19).
Materials and Methods
Brief description of Machine Learning Classification Techniques
Logistic Regression
Logistic Regression (LR) is a sort of supervised learning which estimates the connection between a binary dependent variable and at least one independent variable by evaluating probabilities with the help of sigmoid function. In contrary to its name, LR is not used for regression problems rather is a type of machine learning classification problem where the dependent variable is dichotomous (0/1, -1/1, true/false) and independent variable can binominal, ordinal, interval or ratio-level. The sigmoid/Logistic function is given as:
                                                       (1)
where, y is the output which is the result of weighted sum of input variables x. If the output is more than 0.5, the output is 1 else the output is 0 (19).
K-Nearest Neighbor
K-Nearest Neighbor (KNN) method can be used for both classification and regression problems (19). However, it is more widely used in classification problems in the industry. KNN is a simple algorithm that stores all available cases and classifies new cases by a majority vote of its k neighbors. The case is assigned to the class is most common amongst its K nearest neighbors measured by a distance function. These distance functions can be Euclidean, Manhattan, Minkowski, and Hamming distance.
The first three functions are used for continuous function and the fourth one (Hamming) for categorical variables. If K = 1, then the case is simply assigned to the class of its nearest neighbor. At times, choosing K turns out to be a challenge while performing KNN modeling. Its major advantage is the simplicity of translation and low computation time.
Support Vector Machine
Support Vector Machine (SVM) is a supervised classifier in machine learning algorithms that can be used both for regression and classification. It is majorly applied in solving classification problems. The goal of SVM is to classify data points by an appropriate hyperplane in a multidimensional space. A hyperplane is a decision boundary to classify data points. The hyperplane classifies the data points with the maximum margin between the classes and the hyperplane (19).
In this algorithm, we plot each data item as a point in n-dimensional space (where n is the number of features you have) with the value of each feature being the value of a particular coordinate. For example, if we only had two features like Height and Hair length of an individual, we’d first plot these two variables in two-dimensional space where each point has two coordinates (these coordinates are known as Support Vectors) In fig 1, the line which splits the data into two differently classified groups is the black line since the two closest points are the farthest apart from the line. This line is our classifier. Then, depending on where the testing data lands on either side of the line, that’s what class we can classify the new data as (25).

 

Fig 1. Support vector machine
 

 

Fig 3. A multilayer perceptron with two hidden layers (35)
 

 

Algorithm1 gradient boost

1. Let ƒ0 be a constant 2.  For i = 1 to M            a. Compute gi(x) using eq()            b. Train the function h (x, θi)            c. Find 𝜌i using eq()            d. Update the function                                       𝑓i = 𝑓i-1 + 𝜌i ℎ (𝑥, 𝜃i) 3. End

Prediction
Positive			Negative
Actual	Positive Negative	TP FP	FN     TN

 

Variables Name	Normal Interval	Mean, STD	Type
Age	20-93	48.085057-14.686155	N
Gender	Female or Male	1.351615-0.477530	N
Weight	36-141 (kg)	67.525920-12.443953	N
BMI	15.2-45.8 (kg/m2)	24.123923-3.397294	N
SBP	77-200 (mmHg)	123.219382-17.513858	N
DBP	44-134 (mmHg)	76.360446-11.004056	N
FPG	1.78-6.99 (mmol/L)	5.226368-0.781089	N
FFPG	3.2-29.7 (mmol/L)	5.729084-1.424568	N
Cholesterol	2.12-11.65 (mmol/L)	4.860753-0.925377	N
Triglyceride	0-32.64 (mmol/L)	1.588476-1.260003	N
HDL	0-2.85 (mmol/L)	1.666765-1.079377	N
LDL	0.55-6.75 (U/L)	2.976220-0.892196	N
ALT	0.4-365.7 (mmol/L)	26.754810-22.279922	N
BUN	1.38-17.73 (mmol/L)	4.841088-1.251566	N
CCR	33.9-307 (mmol/L)	71.557080-16.937394	N
Smoking status	1, current smoker 2, ever smoker 3, never smoker	3.006123-1.199571	N
Drinking status	1, current drinker 2, ever drinker 3, never drinker	3.209935-0.939842	N
Family history	0-1	0.061585-0.240428	N
Diabetes	0-1	0.302812-0.459528	N

 

Model selection using k-fold cross validation
In our work, we have carried out parameter tuning, have trained models with optimal parameters, and made predictions using these models by implementing the k-fold cross-validation procedure.
 

 

Algorithm: proposed method for type 2 diabetes prediction

1. begin 2. pre-processing of the dataset different statistical techniques 3. using k-fold cross validation technique for tuning parameters and best model selection 4. classification of diabetes and healthy people using GBM, LR, KNN, NB, DT, RF, SVM, NN classifiers 5. computes different performance evaluation metrics for model evaluation 6. selection best model for type 2 diabetes 7. end

 

The cross-validation steps be summarized as follows:
Step 1. The available dataset is split into two parts, training and an independent test set. The training set is used for parameter tuning and test set for the final model evaluation step at the end i.e. at Step 5. In our work, 70% of the combined dataset is randomly selected from the available dataset to be used as the training set while the remaining 30% is used as the test set. 
Step 2. In the second step, we specify the range of values for the different parameters that are to be tuned. 
Step 3. In this step, for each parameter configuration, the k-fold cross-validation is applied to the training set. The best parameter configuration is identified. we utilized a grid search to identify the best set of parameters of the models generated from different methods. 10-fold stratified cross-validation (SCV) was utilized by us. Accuracy was used as the performance measure to evaluate the parameters. 
Step 4. In this step, models of algorithms are constructed using the optimal parameters obtained in the previous step.
Step 5. In this step, the test data set mentioned in the first step is used to evaluate the models made in step 4. The model that has the best set of evaluation criteria in evaluations. Selected as the best model.
Step 6. Finally, after evaluating the models, after combining the training and test sets, the winning model can be fitted to all data. This model can be used for deployment.
Result and Discussion
Statistical analysis and model validation
In this section, the models are compared with each other to show their performance based on the dataset of China population. High accuracy means that the model predicts the accurate output. By considering accuracy, it could be possible to understand whether the model was trained well or not. Generally, how is its performance? Consequently, in analyzing the results of model accuracy, the gradient boosting model had shown the best performance.  However, this criterion does not give detailed information about model performance. Therefore, there is a need for other criteria in order to analyze further.
To compare the machine learning methods, the diabetic dataset is used and tested by implementing LR, DT, RF, SVM, NB, GBM, KNN, and NN. The training Accuracy, test Accuracy, Sensitivity, Specificity, Precision, F1_Score, MSE, RMSE of performance metrics for ML models, are reported in Table 3 and fig 6. These metrics have been calculated using the test set.

 

ROC	RMSE	MSE	F1	Pr	SP	SE	Test Acc	Trainig Acc	Algorithm
0.93	0.215	0.046	0.916	0.945	0.889	0.979	0.953	0.956	LR
0.94	0.215	0.046	0.917	0.933	0.9	0.973	0.953	0.954	DT
0.93	0.219	0.048	0.913	0.942	0.886	0.978	0.953	1	RF
0.93	0.213	0.045	0.916	0.964	0.873	0.986	0.954	0.956	SVM
0.92	0.252	0.063	0.889	0.889	0.889	0.955	0.936	0.946	NB
0.94	0.211	0.045	0.917	0.924	0.9	0.988	0.955	1	GBM
0.92	0.226	0.051	0.906	0.955	0.862	0.983	0.948	0.964	KNN
0.93	0.236	0.055	0.901	0.911	0.892	0.965	0.944	0.981	NN

 

From the results presented in Table 3, it can be observed that out of the compared algorithms, NB performed the worst on the test set. The NB model possesses the lowest values for all metrics. This is in an agreement where NB performed the worst amongst the ML methods compared to them.
KNN and NN (1_layer perceptron) in our work, performs significantly better than NB. In the same way, LR and RF perform better than the NB, KNN, and NN classifiers and are fourth and Fifth best in terms of the values of metrics.
SVM performs better than the other classifiers and is Third best in terms of the values of the Precision and MSE. The values are 0.964 and 0.045 respectively. DT performs better than the other classifiers and is second best in terms of the values of the SP, F1, and ROC. The values are 0.900, 0.917 and 0.94 respectively.
Among them, GBM possessed the highest performance metrics with Train Acc, Test Acc, SE, SP, F1, MSE, and RMSE being 1, 0.955, 0.988, 0.900, 0.917, 0.045, and 0.211 respectively. Since GBM is scalable, has the potential to serve as a general-purpose algorithm to develop models for diabetes predictor using datasets.
 

Fig 6. Statistical analysis for ML models

 

In terms of area under the receiver operating characteristic curve (ROC) metric, GBM and DT outperformed other compared methods on datasets. GBM and DT maintained the best results with the ROC of 0.94, compared to other ML methods. (see Fig 7)

 

Fig 7. Receiver Operating Characteristic Curve (ROC) Metric

Conclusion
In this research, by applying eight learning machine algorithms, different models for diagnosis of diabetes were made. They were analyzed and compared with each other by using various valuation parameters. The result of the test shows that the gradient boosting model is the best model with an accuracy of 95.50% in the dataset of the Chinese population.
So far, machine learning methods have not been used to predict diabetes in the studied data set. This study could pave the way for others to research this data set. The basis of this study is to do more research and develop models such as other machine learning algorithm.
 

How to cite: Karimi Darabi P, Tarokh M.J. Type 2 Diabetes Prediction Using Machine Learning Algorithm. Jorjani Biomedicine Journal. 2020; 8(3): 4-18.