A Study of Creatinine Level among Patients with Dyslipidemia and Type 2 Diabetes Mellitus using Multilayer Perceptron and Multiple Linear Regression.

BACKGROUND AND
OBJECTIVE: Dyslipidemia is one of the most important risk factors for coronary heart disease with diabetes mellitus. Diabetic dyslipidemia is correlated with reduced concentrations of high-density lipoprotein cholesterol, elevated concentrations of plasma triglycerides, and increased concentrations of dense small particles of low-density lipoprotein cholesterol. Furthermore, dyslipidemia is one of the factors that accelerate renal failure in patients with nephropathy that is observed to be higher in these patients. This paper aims to propose the variable selection using the multilayer perceptron (MLP) neural network methodology before performing the multiple linear regression (MLR) modeling. Dataset consists of patient with Dyslipidemia, and Type 2 Diabetes Mellitus was selected to illustrate the design-build methodology. According to clinical expert's opinion and based on their assessment, these variables were chosen, which comprises the level of creatinine, urea, total cholesterol, uric acid, sodium, and HbA1c.
MATERIALS AND METHODS: At the first stage, all the selected variables will be a screen for their clinical important point of view, and it was found that creatinine has a significant relationship to the level of urea reading, a total of cholesterol reading, and the level of uric acid reading. By considering the level of significance, α = 0.05, these three variables are being selected and used for the input of the MLP model. Then, the MLR is being applied according to the best variable obtained through MLP process.
RESULTS: Through the testing/out-sample mean squared error (MSE), the performance of MLP was assessed. MSE is an indication of the distance from the actual findings from our estimates. The smallest MSE of the MLP shows the best variable selection combination in the model.
CONCLUSION: In this research paper, we also provide the R syntax for MLP better illustration. The key factors associated with creatinine were urea, total cholesterol, and uric acid in patients with dyslipidemia and type 2 diabetes mellitus. Copyright:

INTRODUCTION

Dyslipidemia and type 2 diabetes mellitus

Diabetes mellitus is chronic metabolic disorder, which if left undiagnosed or poorly management poses various risks to patient and comprises the quality of life. The compromised immune state may bring limitations to treatment planning for oral and systemic aliments.[1] Dyslipidemia is one of the major risk factors for coronary heart disease with diabetes mellitus. Diabetic dyslipidemia is associated with decreased high-density lipoprotein cholesterol concentrations, high plasma triglyceride concentrations, and increased concentrations of dense small low-density lipoprotein (LDL) cholesterol particles.[23] The actual pathogenesis of diabetic dyslipidemia is not confessed although much evidence suggests that insulin resilience plays an important role in the development of the disease.[4] The leading cause of diabetic dyslipidemia is that insulin-resistant fat cells release an increase in free fatty acids that induce triglyceride production which eventually induces the secretion of apolipoprotein B and very LDL cholesterol.[5] Epidemiological studies have shown that diabetes mellitus is an independent cardiovascular risk factor and that it intensifies the symptoms of other known risk factors, such as hypertension, smoking, and hypercholesterolemia.[6]

Dyslipidemia is an important problematic medical condition which initiates fatal cardiac diseases. This condition is frequently marked by a higher than the normal value which is hyperlipidemia. Dyslipidemia is one of the factors that accelerate renal failure in nephropathy patients which is found higher in these groups of patients.[7] The renal failure could be related to various other pathologies such as vessel abnormalities or metabolic disorders such as renin impairment. Controlling dyslipidemia in nephropathy could reverse or slow the renal impairments as discussed by various researchers and added to the renal deterioration in functions.[8] On the other hand, creatinine is an important marker for renal impairment. High creatinine is known as associated with renal abnormalities. Interestingly, low serum creatinine, on the other hand, can increase the risk of diabetes mellitus.[9]

This study aims to classify the major factors associated with creatinine in dyslipidemia and type 2 diabetes mellitus patients. During the treatment of dyslipidemia with type 2 diabetes mellitus, this knowledge could be beneficial and the related factors with a disease could be identified that could shed light on the availability of data at Hospital Universiti Sains Malaysia (HUSM).

MATERIALS AND METHODS

Multilayer perceptron neural network

Multilayer perceptron (MLP) procedure will be applied, which is the most widely used artificial neural network. MLP is normally divided into three key layers that consist of the input, the hidden, and the output layer.[101112] In the investigation study, the output node of this analysis is one because only one dependent variable exists. Equation (1) gives the MLP with N input nodes, H hidden nodes, and one output node. The value is shown as follows:

Where w is an output weight from hidden node j to the output node, is the bias for the output node, and g is an activation function. The values of the hidden node h, j = 1 … H are given by

Where v is the output weight from input node i to has hidden node j, v is the bias for hidden node j where j = 1 … H, and x are the independent variables where i = 1 … N and k is an activation function.[101112] Figure 1 displays the general MLP model architecture.

Figure 1

The general architecture of the multilayer perceptron with one hidden layer, N input nodes, H hidden nodes, and one output node

The selected variable from the MLP procedure will be the input for the multiple linear regression (MLR).[13] MLR to provide more than one explanatory variable extends simple linear regression. The proposed model is given as follows:

Creatinine = β0 + β1 Urea Reading + β2 Total Cholesterol + β3 Urid Acid + ε

Where B0, β1, β2 and β3 are regression coefficients, urea reading is referring to urea measurement, total cholesterol is referring to total cholesterol measurement, urid acid is referring to the uric acid measurement, ε is random error, ε∼N(0,σ2).

Data and the R syntax

We used a research data collection of patients with underlying type 2 diabetes mellitus with dyslipidemia disease visiting the Hospital USM outpatient clinic. A total of 30 patients took part in this study. The data summary for the selected variable in the analysis is described in Table 1.

Table 1

Data description of the selected variable in the study

n	Code-variables	Explanation of user variables
1	Y, creatinine (creat)	Creatinine reading
2	X1, urea reading (urea)	Urea reading measurement
3	X2, total cholesterol (TC)	Total cholesterol reading
4	X3, uric acid (uric)	Uric acid reading

R syntax for the multilayer perceptron methodology and multiple linear regression

Input =(”

TC Urea Creat Uric

1.96 5.70 97.00 419.00

6.04 5.20 129.00 373.00

4.93 5.20 83.00 445.00

5.79 5.60 124.00 382.00

3.40 5.70 111.00 357.00

5.62 4.20 113.00 497.00

4.95 4.60 99.00 353.00

3.07 7.00 87.00 438.00

4.02 7.50 125.00 607.00

3.80 8.00 123.00 565.00

3.81 5.20 94.00 413.00

5.01 5.60 101.00 304.00

4.35 7.50 149.00 567.00

3.56 3.90 98.00 350.00

4.20 4.70 107.00 336.00

3.94 5.30 106.00 398.00

6.04 2.90 63.00 233.00

1.45 5.80 64.00 362.00

7.18 4.60 85.00 237.00

4.71 5.70 94.00 424.00

3.74 6.30 121.00 353.00

3.88 3.20 77.00 243.00

4.03 4.10 106.00 329.00

5.42 5.30 133.00 340.00

4.59 4.70 66.00 246.00

3.70 4.50 91.00 492.00

5.43 10.10 168.00 589.00

4.72 6.60 152.00 636.00

5.31 8.90 150.00 480.00

4.54 4.30 91.00 359.00

”)

data = read-table (textConnection (Input), header = TRUE)

mydata<- rbind.data-frame (data, stringsAsFactors = FALSE)

iboot<- sample (1:nrow (mydata), size = 10, replace = TRUE)

bootdata<- mydata[iboot,]

if(!require (neuralnet)){install.packages(”neuralnet”)}

library(”neuralnet”)

apply (bootdata, 2, function (x) sum (is.na (x)))

max_data<- apply (bootdata, 2, max)

min_data<- apply (bootdata, 2, min)

data_scaled<- scale (bootdata, center = min_data, scale = max_data - min_data)

index = sample (1:nrow (bootdata), round (0.70*nrow (bootdata)))

train_data<- as.data-frame (data_scaled[index,])

test_data<- as.data.frame (data_scaled[-index,])

print (train_data)

print (test_data )

n = names (bootdata)

f = as.formula (paste(”Creat ~”, paste (n[!n %in% “Creat”], collapse = “ + “)))

nn = neuralnet (f, data = train_data, hidden = c (2), linear.output = T)

plot (nn)

options (warn=-1)

predicted<- compute (nn, test_data[,1:3])

MSE.net <- sum((test_data$Creat - predicted$net.result)^2)/nrow (test_data)

MSE.net

Model <- lm (Creat ~ TC + Urea + Uric, data = data)

summary (Model)

test<- data[-index,]

predict_lm<- predict (Model, test)

MSE.lm<- sum((predict_lm - test$Creat)^2)/nrow (test)

MSE.lm

print (paste (MSE.lm, MSE.net))

In this case, there are three selected variables, which define as X1 (urea reading), X2 (total cholesterol), and X3 (uric acid). All the selected variables were tested using MLP, and the most significant variable will be used for the regression modeling. In this study, the dataset was partitioned into a training set of 70% and a testing set of 30%. One-hidden-layer MLP is found to be the most suitable model for the studied case.

RESULTS

Figure 2 shows the architecture of the MLP with three input nodes, one hidden layer with two neurons and one output node. In this section, the variable selection had been determining using the developed MLP methodology. Three factors which are urea reading, total cholesterol, and uric acid reading have significantly influenced the creatinine level. The purpose of this present research is the study of the performance of the MLP neural network and MLR. The combination of selected variables that produces the smallest MSE will be considered as the best model for MLP. This result will be obtained by listing and taking all the MSE of the conducted MLP, with a different combination of the variables.

Figure 2

The architecture of the multilayer perceptron with three input nodes, one hidden layer with two neurons and one output node

The most significant variable (by looking at the smallest MSE) which influences the level of creatinine will be considered as the input for the model of MLR. This study proposed the MLP, which consists of three input nodes, one hidden layer and one single output. The output node is set at one, creatinine level (a dependent variable) in this study. The train to test split is 70:30; 70% of the data available for network training and the remaining 30% for network testing.[101114] MLP performance was measured by the testing/out-sample MSE. MSE shows how far our estimates vary from the actual results. Table 2 summarizes the result for the multiple regression modeling (MLR). The models are shown below.

Table 2

Coefficients result of multiple regression

Model	Unstandardized coefficients		t	Significant	95% CI for B
	B	SE			Lower bound	Upper bound
Constant	−6.75	17.20	−0.39	0.69	−42.10	28.61
Urea reading	7.54	2.67	2.82	0.01	2.06	13.03
Total cholesterol	7.46	2.520	2.96	0.01	2.28	12.65
Uric acid	0.09	0.04	−0.392	0.02	−42.10	28.61

Dependent variable: Creatinine. R2: 0.746, (F [df]=18.06 [3, 26]; P<0.05). Multiple linear regression. Model Assumption is met. CI: Confidence interval, SE: Standard error

Therefore the proposed linear model is given by

Creatinine = −6.75 + 7.54 (urea reading) + 7.46 (total cholesterol) + 0.09 (urid acid) (2).

Equation (2) gives the multiple linear models of the creatinine level. The urea reading (β1: 7.54; P < 0.05; 95% confidence interval [CI] [2.06, 13.03]) shows a significant relationship toward the creatinine level. Total cholesterol reading (β2: 7.46; P < 0.05; 95% CI [2.28, 12.65]) also shows a significant relationship to the level of creatinine. The third variable is the acid uric reading (β3: 7.46; P < 0.05; 95% CI (−4110, 28.61)).

DISCUSSION

The key emphasis of this paper is the development of methodologies for MLP and MLR. At the first step, data were divided into the training dataset 70% and testing dataset 30%. The MLP model at one hidden layer was applied. The data were obtained from the Unit of Record, Hospital USM. In this paper, we calculate the mean MSE of MLP and MLR. This is to assess the performance of the network; at the same time, it can be used as the variable selection procedure. The smallest MSE obtained from the possible MLP model will be selected. The input of the MLP will be used for the MLR model building. This is to ensure that the model obtains will be the best model for the prediction purposed. Using these methods, it is shown that urea, total cholesterol, and uric acid have significantly influenced the creatinine level. Creatinine levels are considered to be a significant predictor of kidney disease.

CONCLUSION

The presence of these three factors, urea, total cholesterol, and uric acid, would enhance the level of creatinine. In diabetes, the severity of the diabetes of a patient is influenced by the urea but not creatinine levels.[1516] Total cholesterol is more correlated with glucose level compared to creatinine.[17] This could point out that urea and total cholesterol are closely related to glucose compared to creatinine. Our data are contradicted with Bagheri et al., (2019), which showed a negative correlation between creatinine and total cholesterol.[18] However, these three factors may influence the creatinine in the condition where renal functions have already deteriorated as suggested by other studies mentioned above. This information hopefully will help a clinician to manage these three factors to help to improve the creatinine and renal functions. In conclusion, urea, total cholesterol, and uric acid were essential factors for creatinine in patients with dyslipidemia and type 2 diabetes mellitus.

Financial support and sponsorship

The authors would like to express their gratitude to Universiti Sains Malaysia (USM) for providing the research funding (No. 304/PPSG/6315410, School of Dental Sciences [PPSG], Health Campus, USM).

Conflicts of interest

There are no conflicts of interest.