Accuracy Improved Classification and Regression Tree (CART) Model: Diabetes Prediction Using Minority Over-Sampling and Particle Swarm Optimization Techniques

2.1. Literature Review

On the research approach and design stage, reference materials such as books, articles, journals and other scientific works will be collected to support the objects and methods used in the research. This reference will be used to improve the accuracy of the CART model in predicting diabetes with the SMOTE and PSO techniques.

2.2. Data and Data Source

The data used is from the kaggle website with the title "Diabetes Health Indicators" which is sourced from the 2015 BRFSS results report consisting of 441,456 rows and 330 attributes of raw data which will later be selected again by selecting the attributes that will be used to retrieve the new dataset "Diabetes Health Indicators" with 21 dependent attributes and 1 independent target used in classifying people with diabetes and not in a person. Attribute data used as follows.

2.3. Data Analysis

The stages of data analysis in this study to apply increased accuracy to the decision tree algorithm with the classification and regression tree (CART) model for the prediction of diabetes using the synthetic minority over-sampling technique (SMOTE) and particle swarm optimization (PSO) are as follows:

Import Data

Import Data is the process of retrieving data files that have been provided from outside the application, namely through the dataset from kaggle "Diabetes Health Indicators Dataset". Data that is ready to be imported into jupyter notebook.

Preprocessing

On the preprocessing stage, it can be used to identify and improve the data to be studied. The dataset that has class imbalance is then oversampled using the SMOTE technique. The equation for SMOTE is based on [20] as follows :

$$x_{syn}=x_i+{(x_{knn}-x_i)}^{}x\tau$$

As:

$x_{syn}$ =data from resampling

$x_i$ =data that will be replicated

$x_{knn}$ =data with the closes distance from the replicated data.

$\tau$ = random numbering 0-1

Feature Selection

Data that has been oversampled will be carried out to a feature selection or attribute selection using PSO. In solving optimization problems and relevant feature selection problems, this is one of the benefits of the PSO algorithm. The feature selection process uses the PSO algorithm according to [21], the steps are as follow:

• Initialization: done randomly to determine the initial particle.
• Fitness: a measure in each particle in the population.
• Update: calculate the velocity of each particle with the equation below.

$$v_j^{t+1}=w.v_j^t+c_1.r_1\left({pBest}_j^t-x_j^t\right)+c_2.r_2\left({gBest}_j^t-x_j^t\right)$$

As:

$v_j^{t+1}$ = particle velocity $j$ on iteration

$v_j^t$ = particle velocity $j$ on iteration to- $t$$x_j^t$ = particle position $j$ on iteration to- $t$$w$ = inertial weight

$c_1,c_2$ = constant velocity

$r_1,r_2$ = random numbering $\in \left[\mathrm{0,1}\right]$${pBest}_j^t$ = best position for particle $j$ on iteration to- $t$${gBest}_j^t$ = global optimal for particle $j$ on iteration to- $t$

1)

Construction with the equation below.

$$x_j^{t+1}=x_j^t+v_j^{t+1}$$

2)

Termination: stop the process, if the termination criteria are met, and return to step 2 (fitness) if not met.

Split Data

On the data splitting process, the dataset is divided into two parts. Namely, test data and training data with a ratio of 80:20, 80% as training data and 20% as test data.

Algorithm clarification method on decision tree model Classification and Regression Trees (CART)

On the CART classification process with stages:

1)

Determination of the variables to be tested.

2)

Determining the number of sorters per variable according to the type of independent variable using the equation below.

$b-1selection$ : Continues independent variable

$2^{L-1}-1selection$ : Nominal category independent variable

$L-1pemilahan$ : Ordinal category independent variable

As:

$b$ = Amount of data on a variable

$L$ = Amount of category on a variable

3)

Calculate the ‘Gini index’ value for each sorter according to the equation below.

$$i\left(t\right)=1-{\sum }_{j=1}{P}^2\left(j\right|t)$$

As:

$i\left(t\right)$ = Gini index

$P\left(j|t\right)$ = Class proportion j on node t

$n_j\left(t\right)$ = Amount of observation on class j node t

$n\left(t\right)$ = Amount of observation on node t

Then the sorter that has the smallest Gini index value will be chosen to be the best sorter.

4)

Repeat step 2-3 to perform sorting until it is no longer possible.

5)

Marking of terminal node class labels is based on the rules for the highest number of members using the equation below.

$$P\left(j|t\right)={max}_j\frac{N_j\left(t\right)}{N\left(t\right)}$$

As:

$P\left(j|t\right)$ = Probability of class j on node t

$N_j\left(t\right)$ = Amount of observation on class j node t

$N\left(t\right)$ = Amount of observation on node t

Testing and Algorithm Evaluation

Testing and evaluation process functions as a model test and calculates the accuracy produced by the Decision Tree model Classification and Regression Trees (CART) algorithm. The testing process applies k-fold cross validation with a k-fold value = 10 and will be evaluated using a confusion matrix. The evaluation stage uses the confusion matrix as follows.

1)

Insert testing result on the confusion matrix Table 1.

As:

TN : True Negatif

FP : False Positive

FN : False Negatif

TP : True Positive

2)

Calculate the accuracy value using the equation.

$$\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{r}\mathrm{y}=\frac{TP+TN}{TP+FP+TN+FN}$$

The level of precission using the equation.

$$\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{TP}{TP+FP}$$

Recall value using the equation.

$$\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{TP}{TP+FN}$$

The value of the F1 score uses the equation.

$$\mathrm{F}1\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\frac{2\times \mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$$

APER value uses the equation.

$$APER=\frac{FP+FN}{TP+FP+TN+FN}x100\%$$

MAE (Mean Absolute Error) value is used to calculate the average difference between the calculated value and the actual value [22] uses the equation.

$$MAE=\frac{\sum _{\{u,i\}}\left|p_{ui}-r_{ui}\right|}{N}$$

As:

$p_{ui}$ = prediction value.

$r_{ui}$ = actual value.

N = amount of data lines.

4.1. Data Mining

Data mined and used in this research is "Diabetes Health Indicators Dataset" from Kaggle in the form of Comma Separated Values ​​(CSV). The attributes used are 22 attributes namely.

Table 2. Data attribute on Diabetes Health Indicators Dataset.

No	Attribute	Variable Name
1	Diabetes Disease	DIABETE3
2	High Blood Pressure	_RFHYPE5
3	High Cholesterol	TOLDHI2
4	Cholesterol Check	_CHOLCHK
5	BMI (Body Mass Index)	_BMI5
6	Smoker	SMOKE100
7	Physical Activity	_TOTINDA
8	Consume Fruit	_FRTLT1
9	Consume Vegetables	_VEGLT1
10	Alcohol Consumption	_RFDRHV5
11	Stroke	CVDSTRK3
12	Heart Disease	_MICHD
13	Health Care Coverage	HLTHPLN1
14	Medical Checkup	MEDCOST
15	General Health	GENHLTH
16	Mental Health	MENTHLTH
17	Physical Health	PHYSHLTH
18	Difficult Walking Or Climbing Stairs	DIFFWALK
19	Sex	SEX
20	Age Category	_AGEG5YR
21	Highest Grade School	EDUCA
22	Income	INCOME2

Table 2. Data attribute on Diabetes Health Indicators Dataset.

No	Attribute	Variable Name
1	Diabetes Disease	DIABETE3
2	High Blood Pressure	_RFHYPE5
3	High Cholesterol	TOLDHI2
4	Cholesterol Check	_CHOLCHK
5	BMI (Body Mass Index)	_BMI5
6	Smoker	SMOKE100
7	Physical Activity	_TOTINDA
8	Consume Fruit	_FRTLT1
9	Consume Vegetables	_VEGLT1
10	Alcohol Consumption	_RFDRHV5
11	Stroke	CVDSTRK3
12	Heart Disease	_MICHD
13	Health Care Coverage	HLTHPLN1
14	Medical Checkup	MEDCOST
15	General Health	GENHLTH
16	Mental Health	MENTHLTH
17	Physical Health	PHYSHLTH
18	Difficult Walking Or Climbing Stairs	DIFFWALK
19	Sex	SEX
20	Age Category	_AGEG5YR
21	Highest Grade School	EDUCA
22	Income	INCOME2

4.2. Data Cleaning

The initial process carried out in data preprocessing is cleaning data which aims to remove empty data or missing values. Check whether there is a missing value or not in each attribute using the isnull() and sum() functions. Produces the missing values ​​shown in Table 3.

After that, the drop missing value was carried out using the dropna() function to produce 343,406 new data rows and 22 attributes.

4.3. Formatting Data

The data formatting stage aims to standardize the format of the dataset used in the research. The formatting performed on the Diabetes Health Indicators Dataset is shown in Table 4.

After the data formatting process for each attribute, the attribute name is changed with the rename(columns) function which aims to make it easier to read each attribute in the way shown in Figure 1.

Output result of dataset diabetes_health_indicator:

Figure 2. Output result for upper part of diabetes health indicator dataset.

Figure 2. Output result for upper part of diabetes health indicator dataset.

4.4. Oversampling Data Using SMOTE Technique

The process that is carried out before oversampling is to ensure that the data that is owned is not similar to the other data. Using the duplicate() function, there are 23960 rows that have similarities between rows. Data that has similarities is done by dropping duplicates() so that it produces duplicate data to 0. Separation of data for the independent variable and the dependent variable with ‘y’ as the dependent variable, namely the attribute of Diabetes and ‘x’ as the independent variable, namely the attribute other than Diabetes.

The diabetes attribute has class data that is not balanced with the number of those with diabetes totaling 39657 data and those who are not affected by diabetes totaling 190055 data shown in Figure 3. The results of oversampling on the diabetes attribute with the SMOTE technique using the library imblearn.over_sampling class data that is not balanced with the number those with diabetes have an increase of 65.47% in the class of data affected by diabetes totaling 190055 data and those not affected by diabetes still amounting to 190055 data, shown in Figure 4.

4.5. Feature Selection Algoritma PSO

In the feature selection process using the PSO algorithm, it is done by determining several parameters. Parameter determination is based on research that has been conducted by [23] the parameter includes the constant velocity (${\mathit{C}}_1,{\mathit{C}}_2$) with the value of 1.49, inertia weight (w) with the value of 0.72, amount of iteration or repeating (N) as much as 100 times and amount of parameter for particles as much as 30. After 100 iterations, the attributes generated by the PSO algorithm will be obtained with a threshold value used to select the best feature with a score above 0.5.

The PSO process in python using the niapy, sys and sklearn libraries shows the results of the selected attributes as many as 7 attributes, namely the attribute HighBP, HighChol, Smoker, Fruits, NoDocbcCost, GenHlth and Age.

Figure 5. The selected attribute results from the PSO algorithm.

Figure 5. The selected attribute results from the PSO algorithm.

The selected attributes have a correlation value between the attributes shown in Figure 6.

The correlation value between the dependent attribute, namely diabetes, and the attribute selected by the PSO algorithm is shown in Table 5.

4.6. Data Mining Result

At the data mining stage, there are two mining processes. First, the classification process with the CART Model Decision Tree algorithm on the Diabetes Health Indicators Dataset. Second, the classification process uses the CART Decision Tree algorithm on the Diabetes Health Indicators Dataset with data oversampling using the SMOTE technique and feature selection using the PSO algorithm.

Data mining with the CART model decision tree algorithm uses the skleran library with import DecisionTreeClassifer and input category 'gini' for classification using the gini index. To evaluate the model using the sklearn.model_selection library with import cross_val_score, KFold and train_test_split. To evaluate the model based on the matrix from the results of the model evaluation using the sklearn.metrics library with import confusion_matrix and classification_report.

The dataset is divided into two, namely training data and test data with a ratio of 80:20. Then, the training data is processed using the CART algorithm and performed k-fold 10 for model testing. In this study, 3 experiments were carried out and then the performance evaluation of the CART algorithm was calculated using the confusion matrix and the MAE value, from the 3 trials, one of the best average accuracy values ​​and the smallest MAE value would be taken.

 a.

CART Classification Result

The result of the first classification is the application of the CART Model Decision Tree algorithm classification. Diabetes Health Indicators Dataset is classified using the CART Algorithm without using data oversampling and feature selection. The number of attributes used is 21 attributes and 1 target.

• The results of the first CART algorithm trial

Table 6. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the first attempt of CART classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	75,54	31,39	34,12	32,69	24,46	0.2446
2	75,61	30,55	33,39	31,91	24,39	0.2439
3	75,43	30,31	33,70	31,91	24,57	0.2457
4	75,24	30,70	33,51	32,04	24,76	0.2476
5	75,10	30,72	34,01	32,28	24,90	0.2490
6	75,48	30,68	33,53	32,04	24,52	0.2452
7	74,70	30,78	33,72	32,19	25,30	0.2530
8	75,35	31,37	35,71	33,40	24,65	0.2465
9	75,34	29,32	32,31	30,74	24,66	0.2466
10	75,40	29,75	33,79	31,64	24,60	0.2460
Median	75,32	30,56	33,78	32,08	24,68	0.2468

Table 6. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the first attempt of CART classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	75,54	31,39	34,12	32,69	24,46	0.2446
2	75,61	30,55	33,39	31,91	24,39	0.2439
3	75,43	30,31	33,70	31,91	24,57	0.2457
4	75,24	30,70	33,51	32,04	24,76	0.2476
5	75,10	30,72	34,01	32,28	24,90	0.2490
6	75,48	30,68	33,53	32,04	24,52	0.2452
7	74,70	30,78	33,72	32,19	25,30	0.2530
8	75,35	31,37	35,71	33,40	24,65	0.2465
9	75,34	29,32	32,31	30,74	24,66	0.2466
10	75,40	29,75	33,79	31,64	24,60	0.2460
Median	75,32	30,56	33,78	32,08	24,68	0.2468

2.

The results of the second CART algorithms trial

Table 7. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the second trial of CART classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	75,58	31,54	34,34	32,88	24,42	0.2442
2	75,82	30,95	33,52	32,18	24,18	0.2418
3	75,32	30,04	33,42	31,64	24,68	0.2468
4	75,35	30,83	33,38	32,06	24,65	0.2465
5	75,26	31,10	34,38	32,66	24,74	0.2474
6	75,21	30,00	32,82	31,34	24,79	0.2479
7	74,71	30,70	33,46	32,02	25,29	0.2529
8	75,33	31,15	35,11	33,01	24,67	0.2467
9	75,38	29,60	32,90	31,16	24,62	0.2462
10	75,60	30,13	33,97	31,94	24,40	0.2440
Median	75,34	30,58	33,75	32,09	24,66	0.2466

Table 7. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the second trial of CART classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	75,58	31,54	34,34	32,88	24,42	0.2442
2	75,82	30,95	33,52	32,18	24,18	0.2418
3	75,32	30,04	33,42	31,64	24,68	0.2468
4	75,35	30,83	33,38	32,06	24,65	0.2465
5	75,26	31,10	34,38	32,66	24,74	0.2474
6	75,21	30,00	32,82	31,34	24,79	0.2479
7	74,71	30,70	33,46	32,02	25,29	0.2529
8	75,33	31,15	35,11	33,01	24,67	0.2467
9	75,38	29,60	32,90	31,16	24,62	0.2462
10	75,60	30,13	33,97	31,94	24,40	0.2440
Median	75,34	30,58	33,75	32,09	24,66	0.2466

3.

The results of the third CART algorithm trial.

Table 8. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the three CART classification trials.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	75,41	31,10	33,87	32,42	24,59	0.2459
2	75,58	30,73	34,00	32,28	24,42	0.2442
3	75,33	30,14	33,67	31,81	24,67	0.2467
4	75,22	30,49	33,01	31,70	24,78	0.2478
5	75,02	30,64	34,16	32,30	24,98	0.2498
6	75,50	30,81	33,78	32,23	24,50	0.2450
7	74,77	30,88	33,70	32,23	25,23	0.2523
8	75,63	31,83	35,76	33,68	24,37	0.2437
9	75,27	29,34	32,64	30,90	24,73	0.2473
10	75,40	29,82	34,00	31,77	24,60	0.2460
Median	75,33	30,58	33,79	32,10	24,67	0.2467

Table 8. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the three CART classification trials.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	75,41	31,10	33,87	32,42	24,59	0.2459
2	75,58	30,73	34,00	32,28	24,42	0.2442
3	75,33	30,14	33,67	31,81	24,67	0.2467
4	75,22	30,49	33,01	31,70	24,78	0.2478
5	75,02	30,64	34,16	32,30	24,98	0.2498
6	75,50	30,81	33,78	32,23	24,50	0.2450
7	74,77	30,88	33,70	32,23	25,23	0.2523
8	75,63	31,83	35,76	33,68	24,37	0.2437
9	75,27	29,34	32,64	30,90	24,73	0.2473
10	75,40	29,82	34,00	31,77	24,60	0.2460
Median	75,33	30,58	33,79	32,10	24,67	0.2467

The results of the CART algorithm experiment using 3 trials resulted in the best trial in the 2nd trial which is shown in Table 9.

The results of the second experiment showed that the highest accuracy of applying the CART algorithm without using data oversampling and feature selection obtained an average accuracy of 75.34%, precision of 30.58%, recall of 33.75%, f1-score of 32, 09%, APER of 24.66% and an MAE value of 0.2466.

 b.

SMOTE+PSO+CART Classification Results

The result of the second classification is the application of the CART Model Decision Tree algorithm classification. Diabetes Health Indicators Dataset is classified using the CART + SMOTE + PSO Algorithm The number of attributes used is 7 attributes and 1 target.

• The results of the first CART+SMOTE+PSO algorithms.

Table 10. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the first attempt of CART + SMOTE + PSO classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	86,09	92,33	78,73	84,99	13,91	0.1391
2	86,10	92,17	78,91	85,03	13,90	0.1390
3	86,37	93,14	78,74	85,34	13,63	0.1363
4	86,31	92,50	78,94	85,18	13,69	0.1369
5	86,09	91,98	79,00	85,00	13,91	0.1391
6	86,32	92,74	78,93	85,28	13,68	0.1368
7	86,26	92,66	78,88	85,22	13,74	0.1374
8	86,50	92,99	78,99	85,42	13,50	0.1350
9	86,17	92,75	78,52	85,05	13,83	0.1383
10	86,59	92,17	79,56	85,40	13,40	0.1341
Median	86,28	92,54	78,92	85,19	13,72	0.1372

Table 10. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the first attempt of CART + SMOTE + PSO classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	86,09	92,33	78,73	84,99	13,91	0.1391
2	86,10	92,17	78,91	85,03	13,90	0.1390
3	86,37	93,14	78,74	85,34	13,63	0.1363
4	86,31	92,50	78,94	85,18	13,69	0.1369
5	86,09	91,98	79,00	85,00	13,91	0.1391
6	86,32	92,74	78,93	85,28	13,68	0.1368
7	86,26	92,66	78,88	85,22	13,74	0.1374
8	86,50	92,99	78,99	85,42	13,50	0.1350
9	86,17	92,75	78,52	85,05	13,83	0.1383
10	86,59	92,17	79,56	85,40	13,40	0.1341
Median	86,28	92,54	78,92	85,19	13,72	0.1372

2.

The results of the second CART+SMOTE+PSO algorithms.

Table 11. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the second experiment of CART + SMOTE + PSO classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	86,09	92,33	78,73	84,99	13,91	0.1391
2	86,10	92,17	78,91	85,03	13,90	0.1390
3	86,37	93,14	78,74	85,34	13,63	0.1363
4	86,31	92,50	78,94	85,18	13,69	0.1369
5	86,09	91,98	79,00	85,00	13,91	0.1391
6	86,32	92,74	78,93	85,28	13,68	0.1368
7	86,26	92,66	78,88	85,22	13,73	0.1373
8	86,50	92,99	78,99	85,42	13,50	0.1350
9	86,17	92,75	78,52	85,05	13,83	0.1383
10	86,59	92,17	79,56	85,40	13,41	0.1341
Median	86,28	92,54	78,92	85,19	13,72	0.1372

Table 11. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the second experiment of CART + SMOTE + PSO classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	86,09	92,33	78,73	84,99	13,91	0.1391
2	86,10	92,17	78,91	85,03	13,90	0.1390
3	86,37	93,14	78,74	85,34	13,63	0.1363
4	86,31	92,50	78,94	85,18	13,69	0.1369
5	86,09	91,98	79,00	85,00	13,91	0.1391
6	86,32	92,74	78,93	85,28	13,68	0.1368
7	86,26	92,66	78,88	85,22	13,73	0.1373
8	86,50	92,99	78,99	85,42	13,50	0.1350
9	86,17	92,75	78,52	85,05	13,83	0.1383
10	86,59	92,17	79,56	85,40	13,41	0.1341
Median	86,28	92,54	78,92	85,19	13,72	0.1372

3.

The results of the third CART+SMOTE+PSO algorithms.

Table 12. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the second experiment of CART + SMOTE + PSO classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	86,09	92,33	78,73	84,99	13,91	0.1391
2	86,10	92,17	78,91	85,03	13,90	0.1390
3	86,37	93,14	78,74	85,34	13,63	0.1363
4	86,31	92,50	78,94	85,18	13,69	0.1369
5	86,09	91,98	79,00	85,00	13,91	0.1391
6	86,32	92,74	78,93	85,28	13,68	0.1368
7	86,26	92,66	78,88	85,22	13,73	0.1373
8	86,50	92,99	78,99	85,42	13,50	0.1350
9	86,17	92,75	78,52	85,05	13,83	0.1383
10	86,59	92,17	79,56	85,40	13,41	0.1341
Median	86,28	92,54	78,92	85,19	13,72	0.1372

Table 12. The results of each fold for accurary, precision, recall, f1-score, APER and MAE values ​​in the second experiment of CART + SMOTE + PSO classification.

K-fold	Accurary (%)	Precission (%)	Recall (%)	F1-Score (%)	APER (%)	MAE
1	86,09	92,33	78,73	84,99	13,91	0.1391
2	86,10	92,17	78,91	85,03	13,90	0.1390
3	86,37	93,14	78,74	85,34	13,63	0.1363
4	86,31	92,50	78,94	85,18	13,69	0.1369
5	86,09	91,98	79,00	85,00	13,91	0.1391
6	86,32	92,74	78,93	85,28	13,68	0.1368
7	86,26	92,66	78,88	85,22	13,73	0.1373
8	86,50	92,99	78,99	85,42	13,50	0.1350
9	86,17	92,75	78,52	85,05	13,83	0.1383
10	86,59	92,17	79,56	85,40	13,41	0.1341
Median	86,28	92,54	78,92	85,19	13,72	0.1372

The results of the CART algorithm experiment using 3 trials produced the same results from the accurary, precision, recall, f1-score, APER and MAE values ​​shown in Table 13.

The results of the classification experiment show that the highest accuracy of applying the CART algorithm using data oversampling with the SMOTE technique and feature selection with the PSO technique obtains an average accuracy of 86.28%, precision of 92.54%, recall of 78.92%, f1 -score of 85.19%, APER of 13.72% and value MAE 0,1372.

Based on the results above, explained as follows.

 1.

Oversampling Data with SMOTE technique for Diabetes Health Indicators Dataset

On the diabetes attribute, the number of people affected by diabetes is 39,657 data and 190055 data were not affected by diabetes after oversampling the data using the SMOTE technique which results in class balanced data between the number of those affected by diabetes 190055 data and 190055 data that are not affected by diabetes have an increase of 65.47% in the data class affected by diabetes.

 2.

Feature Selection on the Diabetes Health Indicators Dataset attribute using the PSO algorithm

The application of the PSO algorithm for feature selection uses parameters including speed constants ($C_1,C_2$) with a value of 1.49, inertial weight (w) with a value of 0.72, the number of iterations or repetitions (N) is 100 times and the number of parameters for particles is 30 particles, the best feature with a score above 0.5 will be selected by the PSO algorithm. The results of the feature selection attribute were selected as many as 7 attributes, namely the attributes HighBP, HighChol, Smoker, Fruits, NoDocbcCost, GenHlth and Age.

 3.

Accuracy of the CART model Decision Tree algorithm using SMOTE and PSO

By using data oversampling using the SMOTE technique and feature selection with the PSO algorithm to obtain the best accuracy in the CART model decision tree algorithm. Diabetes Health Indicators Dataset has 380110 rows with 7 attributes and 1 target obtaining an average accuracy of 86.28%, precision of 92.54%, recall of 78.92%, f1-score of 85.19%, APER of 13 .72% and the MAE value is 0.1372. Comparison of the accuracy results performed with the CART model decision tree algorithm using and without oversampling and feature selection can be seen in Table 14.