Status of glycosylated hemoglobin and prediction of glycemic control among patients with insulin-treated type 2 diabetes in North China: a multicenter observational study.

BACKGROUND: Blood glucose control is closely related to type 2 diabetes mellitus (T2DM) prognosis. This multicenter study aimed to investigate blood glucose control among patients with insulin-treated T2DM in North China and explore the application value of combining an elastic network (EN) with a machine-learning algorithm to predict glycemic control.
METHODS: Basic information, biochemical indices, and diabetes-related data were collected via questionnaire from 2787 consecutive participants recruited from 27 centers in six cities between January 2016 and December 2017. An EN regression was used to address variable collinearity. Then, three common machine learning algorithms (random forest [RF], support vector machine [SVM], and back propagation artificial neural network [BP-ANN]) were used to simulate and predict blood glucose status. Additionally, a stepwise logistic regression was performed to compare the machine learning models.
RESULTS: The well-controlled blood glucose rate was 45.82% in North China. The multivariable analysis found that hypertension history, atherosclerotic cardiovascular disease history, exercise, and total cholesterol were protective factors in glycosylated hemoglobin (HbA1c) control, while central adiposity, family history, T2DM duration, complications, insulin dose, blood pressure, and hypertension were risk factors for elevated HbA1c. Before the dimensional reduction in the EN, the areas under the curve of RF, SVM, and BP were 0.73, 0.61, and 0.70, respectively, while these figures increased to 0.75, 0.72, and 0.72, respectively, after dimensional reduction. Moreover, the EN and machine learning models had higher sensitivity and accuracy than the logistic regression models (the sensitivity and accuracy of logistic were 0.52 and 0.56; RF: 0.79, 0.70; SVM: 0.84, 0.73; BP-ANN: 0.78, 0.73, respectively).
CONCLUSIONS: More than half of T2DM patients in North China had poor glycemic control and were at a higher risk of developing diabetic complications. The EN and machine learning algorithms are alternative choices, in addition to the traditional logistic model, for building predictive models of blood glucose control in patients with T2DM.

Introduction

It was estimated that in 2017, there were 451 million people with diabetes worldwide; moreover, this figure is expected to increase to 693 million by 2045.[ According to the report of the International Diabetes Federation in 2017, the type 2 diabetes mellitus (T2DM) epidemic in China is the largest in the world, with 11.44 million people living with diabetes, which was expected to increase to 11.98 million by 2045.[ The healthcare expenditure due to T2DM was USD 630 billion in 2017 in China.[ Despite strong financial support, the blood glucose control rate among individuals with diabetes in China is still very low.[

The prognosis of diabetes is mainly referred to based on diabetic complications and comorbidities, including hypertension, cardiovascular diseases, neuropathy, nephropathy, and retinopathy, which can result in poor quality of life and reduced life expectancy.[ However, the World Health Organization has endorsed the use of glycosylated hemoglobin (HbA1c) as a screening test for people at high risk of diabetes and, more importantly, as a test for predicting the risk of microvascular complications.[ Hence, as the gold standard measurement for long-term blood glucose control, HbA1c is a good index of the prognosis of diabetes.

Previous studies have examined the incidence of diabetes and blood glucose control in the Chinese population. For example, Xu et al[ showed that the overall blood glucose control level of the Chinese population in 2010 was 39.7%, but diabetes is related to lifestyle and economic development. Therefore, the data may have changed. Moreover, an increasing number of studies have focused on diabetes in North China. Gao et al[ found that sedentary civil servants had high rates of diabetes in Xinjiang, China. A multicenter study in Shanxi, China described the simultaneous control rate of blood pressure, blood sugar and blood lipids among drug-treated T2DM patients.[ Unfortunately, most of these studies focused on the situation in a single region or special population that rarely touch upon the general situation in North China.[

In recent years, with the development of machine learning, random forest (RF), support vector machine (SVM), and back propagation artificial neural network (BP-ANN) have been widely used in the field of diabetes prediction.[ In addition, studies have shown that combining machine learning algorithms with dimensionality reduction can improve their performance.[ However, few studies have introduced the method of combining dimensionality reduction with machine learning in the field of diabetes.[

Hence, the primary objective of this study was to investigate the general status of diabetic blood glucose control and its influencing factors in insulin-treated T2DM outpatients in North China using multicenter data from various provinces and cities. Moreover, this study explored the application value of combining an elastic network (EN) with a machine-learning algorithm to predict diabetic blood glucose control.

Methods

Ethical approval

The study was approved by the Ethics Committees of Tianjin Medical University (No. TMUHMEC2013032). All patients gave their informed consent.

Subjects

According to the pre-calculated sample size [Supplementary File 1], 27 centers (including secondary hospitals and tertiary hospitals) from six cities in China (Tianjin, Tangshan, Datong, Qinhuangdao, Cangzhou, and Taiyuan) that agreed to participate in the experiment were selected in North China [Supplementary File 2]. From January 2016 to December 2017, data from the first five outpatients were collected daily, and 2787 consecutive patients who met the inclusion and exclusion criteria were eventually included. The inclusion criteria were as follows: (1) T2DM diagnosis, (2) ≥18 years old, and (3) basal insulin use ≥3 months. The exclusion criteria were as follows: (1) refusal to sign informed consent, (2) history of drug allergy or allergies, (3) preparation to get pregnant or current pregnancy or lactation, or (4) psychiatric conditions.

Data collection

Basic information (sex, age, smoking status, alcohol consumption, marital status, etc) and diabetes-related information (typical characteristics of diabetes, duration of diabetes, exercise, diet, oral medications, complications, hypoglycemia, insulin injection time and injection dosage, etc) were collected by questionnaire after the subjects were recruited. The physical examination included height, weight, waist circumference, hip circumference, and blood pressure. The laboratory tests included fasting blood glucose (FBG), 1-h blood glucose (1HBG) and 2-h blood glucose (2HBG), which were measured by the glucose oxidase method; HbA1c, which was determined by high-performance liquid chromatography; total cholesterol (TC) and triglycerides (TGs), which were determined by an enzymatic method; and high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C), which were measured by a homogeneous method.

Quality control

Uniform training for all investigators was provided using standardized procedures. The structured interview method was used to conduct the questionnaire survey to avoid differences in the inquiry process. All data from the questionnaire were thoroughly scrutinized, and missing or illogical options were reassessed and modified.

Statistical analysis

Statistical analyses were performed using the statistical software SAS 9.4 (SAS Inc., North Carolina, USA) and R 3.5.1 (https://www.r-project.org/). The continuous variables were expressed as the median (P25, P75). The categorical variables were represented by the frequency and composition ratio. According to the “Guidelines for the Prevention and Treatment of Type 2 Diabetes in China (2017 Edition),”[ HbA1c <7.0% was defined as the standard of glycemic control. The continuous data and categorical data were compared by the Wilcoxon rank-sum test and Chi-square test, respectively, between the controlled and uncontrolled groups.

Multicollinearity usually exists in the data of chronic diseases,[ so we choose an EN with machine learning validation to obtain a smaller prediction error. It is important to note that ENs eliminate certain predictors to avoid overfitting.[ The EN algorithm is a combination of least absolute shrinkage and selection operator (LASSO) and ridge regression. When there are multiple collinear predictors, LASSO selects only one, ignores others, or zeroes some regression coefficients. The ridge method counteracts collinearity and variance inflation by shrinking the regression coefficients towards zero but without reaching zero. EN combines the penalties of the LASSO and ridge approaches.[ This algorithm achieves sparse coefficient estimates by minimizing the sum of squared errors by adding an ℓ 1 and a squared ℓ 2 penalty simultaneously to the coefficient β.[ The EN removes the limitation of the number of selected variables of LASSO and encourages the grouping effect.

In this study, RF, SVM, and BP-ANN were selected to model and predict diabetes-related variables before and after dimensional reduction by EN. The data were divided into training sets and testing sets on a 7:3 scale. The optimal parameters of each model were found using ten-fold cross-validation, and the area under the receiver operating characteristics curve (AUC) was used to select the optimal model. Then, we estimated the performance of the training model by predicting HbA1c control in the test set. Sensitivity, specificity, accuracy, and AUC were selected to compare the predictive performance of each model. Classification accuracy is a common method in pattern recognition and refers to the ratio of the number of correctly classified samples to the total number of samples. Sensitivity and specificity are the statistical indicators of the performance of classification tests. Sensitivity measures the proportion of actual positive cases that were correctly identified. Specificity refers to the proportion of negativities that were correctly identified. Accuracy, sensitivity, and specificity are expressed as follows, where TP means number of positive cases those are correctly classified; TN means number of negative cases those are correctly classified; FP means number of positive cases those are misclassified; and FN means number of negative cases those are misclassified:

For each algorithm, the final result of the predictive performance is taken as the average value of ten cycles. To compare with machine learning models, this study also established a stepwise logistic regression model.

The RF classifier is composed of multiple tree classifiers, in which each classifier is generated by a random vector independent of the input vector, and the input vector is classified by unit voting.[ The RF classifier used in this study was created by creating a tree using randomly selected features or a combination of features on each node. As a supervised classifier, the basic principle of SVM is to map the input vector to a high-dimensional feature space through the pre-selected non-linear relation and find an optimal classification hyperplane in this space to maximize the classification interval between the two classes.[ BP-ANN is a multilayer feedforward neural network trained by an error back-propagation algorithm, which is one of the most widely used neural network models. The learning rule of BP-ANN is to use the steepest descent method and continuously adjust the weight and threshold of the network through the back-propagation algorithm to minimize the error square sum of the network to identify an optimal model.[

Results

General information of the subjects

According to the inclusion and exclusion criteria, 2787 T2DM patients were enrolled in the study, including 1407 males (50.48%) and 1380 females (49.52%) aged from 19 to 91 years old. Almost all participants (98.49%) were married, and only a few were single (1.51%). Table 1 shows the general information of all subjects in the two groups.

Table 1

General information of patients with insulin-treated type 2 diabetes in HbA1c controlled and uncontrolled group.

Only 1277 (45.82%) subjects met the HbA1c control standard. The percentage of central adiposity in the controlled group was 46.28% compared to 57.02% in the uncontrolled group (χ2 = 31.9762, P < 0.0001). Almost 90% of the subjects in the controlled group exercised, but the rate in the uncontrolled group was only 76.49% (χ2 = 77.7556, P < 0.0001).

The disease conditions and drug treatments of the subjects are shown in Table 2. The overall duration of T2DM was 4 (2.0, 7.5) years and 7 (3.0, 10.0) years in the two groups (Z = –10.5209, P < 0.0001). Over 50% of participants had typical disease characteristics (polydipsia, polyphagia, polyuria, and emaciation) in the uncontrolled group compared to 40.33% in the controlled group. In the controlled group, 33.59% of the subjects had complications, while 53.18% had complications in the non-controlled group (χ2 = 107.5936, P < 0.0001). The rate of hypoglycemia was lower in the controlled group than in the uncontrolled group (14.41% and 23.38%, χ2 = 35.7774, P < 0.0001). The time to the initiation of basal insulin between the two groups was significantly different (χ2 = 23.3311, P < 0.0001). Approximately 9.4% of the subjects were not taking oral hypoglycemic drugs (OHA) in the controlled group, compared to only 6.75% in the uncontrolled group (χ2 = 14.1481, P = 0.0068).

Table 2

Diabetes condition and treatment of patients with insulin-treated type 2 diabetes in HbA1c controlled and uncontrolled groups.

Table 3 shows the serum biochemical indices in the two groups. All biochemical indicators were standardized according to the median values. The differences in FBG, 2HBG, blood pressure, TC, total TGs, HDL-C, and LDL-C between the two groups were statistically significant.

Table 3

Serum biochemical indices of patients with insulin-treated type 2 diabetes in HbA1c controlled and uncontrolled group.

Results of dimension reduction by elastic net

Figure 1 shows the process of EN variable filtering. Table 4 shows the results of dimension reduction by elastic net regression. A total of 42 variables with statistical significance in single factor analysis were entered into the model, and then 19 variables were selected. The variance analysis of the model showed that the model is valid [Table 5] (F = 190.55, P < 0.0001). Hypertensive history, atherosclerotic cardiovascular disease history, nocturnal hypoglycemia, exercise, and TC were protective factors for HbA1c control. Central adiposity, family history, duration of T2DM, typical disease characteristics, complications, insulin dose, OHA, FBG, 2HBG, blood pressure, HDL-C, LDL-C, and hypertension were risk factors for HbA1c control.

Figure 1

Elastic network model fitting process. The abscissa represents the variable selection step and the circle represents the fitting process for each step. ☆ means the best criteria value. Adj R-Sq: Adjusted R2; ASE: The average squared error; AIC: Akaike information criterion; AICC: Akaike information criterion based on Kullback-Laible (KL) information improvement; SBC: Schwartz Bayes criterion.

Table 4

Results of dimension reduction by elastic net.

Table 5

Variance analysis of elastic net.

Results of machine learning

Table 6 shows the prediction results of RF, SVM, BP-ANN, and logistic regression. To increase the reliability of the prediction, ten cycles were performed for each model in this study, and the average value of the ten cycles was used for the final predictive performance evaluation index.

Table 6

Comparison of predictive performance.

According to the results, the sensitivity, specificity, accuracy, and AUC of the RF algorithm without EN dimension reduction for variables were 0.80, 0.71, 0.74, and 0.73, respectively. Although the sensitivity of the reduced dimensional model decreased to 0.79, the specificity, accuracy, and AUC were improved overall. After using an EN to reduce the dimensions of the variables, the sensitivity of the SVM algorithm was improved by 37.70%, specificity by 7.94%, accuracy by 17.74%, and AUC by 18.03%, which were the most improved among the three methods. When the RF algorithm was combined with the variables of dimensionality reduction through EN, the predictive performance was generally improved; the sensitivity was increased from 0.75 to 0.78, the specificity from 0.69 to 0.70, the accuracy from 0.71 to 0.73, and the AUC from 0.70 to 0.72. Among the three machine learning algorithms, RF performed better than SVM and BP-ANN in predicting blood glucose control, both before and after the dimensionality reduction. Generally, after variable screening using the EN first, the three machine learning algorithms were used to predict the long-term blood glucose control in T2DM patients, and this method significantly improved the predictive performance of the abovementioned machine learning algorithm.

Compared with the results of traditional logistic regression, although the logistic specificity was slightly higher than that of the EN and machine learning model, the sensitivity and accuracy (sensitivity = 0.51, accuracy = 0.55) were much lower than those of the EN and machine learning model. In summary, the EN and machine learning model is a good alternative to the traditional logistic regression for blood glucose prediction in individuals with T2DM.

Discussion

In this study, a combination of the EN and machine learning algorithms was used to analyze the factors affecting the standard level of blood glucose among individuals with diabetes, and good results were obtained. On the basis of the traditional Logistic regression model, a new idea on prediction of glycemic control was provided.

Blood glucose control is closely related to complications and prognosis,[ so effective blood glucose control is considered the basis of diabetes treatment. The overall glycemic control rate in our study was 45.82%, which was higher than the 32.60% obtained in a study assessing individuals with HbA1c <7%.[ The results showed that central adiposity, family history, duration of diabetes, blood pressure, and hypertension were risk factors for high HbA1c, which was consistent with previous research conclusions.[ Exercise and modifying vegetable oil intake were protective factors for HbA1c control, which is also consistent with previous studies.[

In this study, the dose of insulin and the use of OHA were risk factors for elevated HbA1c, which may be because the patients taking the high dosage of insulin were mostly patients with severe disease or complications, and their blood glucose control level was relatively poor.[ Unlike previous studies,[ TC was a protective factor for HbA1c control, possibly because a decreased serum TC level can produce an elevated serum interleukin-6 level,[ which is an inflammatory protein related to insulin resistance.[

The results of this study showed that the predictive performance of a machine learning algorithm could be improved by using dimensionality reduction to solve the variable collinearity. It was shown that the EN algorithm has high application value in influencing factor analysis in the field of diabetes and can be used to reduce the dimension of a large number of covariates of diabetes. Future studies should apply the approach of this study to the analysis of diabetes and other chronic diseases.

This study also had some limitations. First, the study did not use rigorous random sampling, and the results may be biased. However, various confounding factors were adjusted to make the results more reliable. Second, this study included only outpatients and did not consider community patients, which may lead to an overestimation of glycemic control. However, the diagnosis and treatment of diabetic patients in China are carried out in secondary and tertiary hospitals. Community hospitals have no specialists and are responsible for only daily medications. Therefore, the research object has a certain representativeness. Third, the study utilized an observational study design and; therefore, failed to produce a causal conclusion that could provide a scientific basis for predicting glycemic control. In future studies, the sample size can be increased on the basis of this study to further explore the analytical ideas.

Acknowledgement

The authors thank all the medical staff who participated in this study for their contributions to this study.

Funding

This study was supported by grants from the Ministry of Education of the Humanities and Social Science Project (No. 17YJAZH048) and the National Natural Science Foundation of China (No. 81803333).

Conflicts of interest

None.