Bayesian network-based risk factors analysis for diabetic retinopathy: A cross-sectional study

doi:10.21203/rs.3.rs-2468734/v1

Background: Diabetic retinopathy (DR) represents one of the common complications of diabetes mellitus, which not only affects the quality of life of patients, but also poses a huge economic burden to their families and society.We aim to construct a Bayesian networks (BNs) to detect risk factors associated with DR, exploring their complicated network relationships, thus, providing a scientific reference for risk prediction and prevention of DR.

Methods: A questionnaire survey and clinical data were collected from 1745 patients with type 2 diabetes from 4 randomly selected communities in four regions of Shanxi Province, China. Afterwards, Elastic Net was employed to make a feature selection, after which, BNs with MMHC algorithms was constructed to discuss risk factors related to DR and their probabilistic dependencies.

Results: After feature selection, 11 variables were incorporated into BNs construction. BNs suggested that the prior probability of DR stands at 0.155. Besides, BNs showed that systolic blood pressure, fasting blood glucose, glycosylated haemoglobin represent parental nodes for DR, and urea nitrogen and urine protein to creatinine ratio constitute child nodes of DR. When the above parameters leaves at abnormal, the probability of DR reaches at 0.741. Besides, marital status and duration of diabetes are indirectly related to DR. Separated marital status indirectly relates to DR by affecting pressure; duration of diabetes indirectly increases the risk of DR through glycated haemoglobin.

Conclusion: The BNs constructed in this study reveals the complex network linkages between DR and its associated factors, and provides a probabilistic model of the risk for DR, allowing physicians and patients to target risk prediction and to take preventive control of DR complicated by type 2 diabetes.

diabetic retinopathy

Elastic network

Bayesian network

risk factors

As the overall birth rate of the natural population increases and mortality decreases, ageing population is accelerating and the incidence of type 2 diabetes mellitus (T2DM), a common chronic non-communicable disease, is increasing annually(1). Diabetic retinopathy (DR) represents one of the common complications of DM, which not only affects the quality of life of patients, but also poses a huge economic burden to their families and society. Due to the complex pathogenesis of DR, the insidious clinical symptoms at the early stage of the disease and the low prevalence of diagnostic tools, early identification and active prevention of the related factors of DR is still an effective way to reduce the occurrence of DR. Therefore, the analysis and reasoning of the factors associated with DR in T2DM patients can enable T2DM patients and clinicians to take timely and effective preventive and control measures to reduce the occurrence of DR or delay the progression of the disease.

Logistic regression model is usually employed to detect risk factors for DR, which assumes that the variables should be inter-independent and uses odds ratio (OR) to reflect the associations. Yet, it is difficult to achieve among clinical parameters. Second, the models fail to present the complex network relationship between risk factors and DR. Besides, it is unable to identify direct and indirect risk factors between the dependent and independent variables(2, 3). Bayesian networks (BNs) are probabilistic models that can visualize the complex internal relationships between diseases and their risk factors by constructing direct acyclic graphs (DAG) to present the direct and indirect relationships between the variables(4), supporting conditional probability tables (CPT) to reflect the associations between variables. Meanwhile, BNs are not subject to strict statistical assumptions and thus can make up for the defects of traditional logistic regression models(5).

The Constraint based (CB) algorithm and the Scoring and searching (SS) algorithm represent the two basic structural algorithms for BNs(6, 7). The CB algorithm first constructs the network through conditional independence tests, and then determines the direction of some edges through independence tests(6), but the construction of directed edges can be contradictory due to the limitations of its independence tests, etc. The SS algorithm uses a scoring function to determine the direction of directed edges by calculating the local scoring maximum for each variable and obtaining the global optimal solution(7), but its algorithmic structure is usually incomplete(8). Max-Min Hill Climbing (MMHC) algorithm is a hybrid algorithm proposed by Tsamardino et al. that could combine the two structural algorithms of BNs(9). It first constructs an undirected network framework using a CB algorithm to reduce the search space, and then uses a SS algorithm to reduce the search space. The MMHC algorithm combines the two algorithms to compensate for the unreliability of the CB algorithm for higher-order conditional independent tests and the incompleteness of the optimal structure derived from the SS algorithm and to fit more relevant factors better.

Risk factors for DR abound, and if all of them are included in the BNs model construction, the model would become complicated. Also, weakly related risk factors would be responsible for a less accurate model. As such, in this study, Elastic Net was first employed to make a screening of variables. Elastic Net was first proposed by Zou et al., aiming to avoid the phenomenon of the traditional variable screening approach, LASSO, which is prone to over-compression in the selection of regression coefficients. It could retain the stability of the ridge regression regularization, and can avoid multiple solutions(10) when screening a large number of variables with high correlation, while generating feature sparsity.

In this paper, we use the Elastic Net model to screen the variables of the original data and retain the risk factors strongly associated with DR, after which, BNs with the MMHC algorithm was constructed to explore the probability dependency relationship between the related risk factors and DR and shed light on the complex internal association with DR and its related factors, so as to take effective preventive measures against the occurrence of DR.

2.1 Subjects

The study population was a stratified random sample of resident type 2 diabetic patients aged 18 to 75 years old, and were selected as per gender and urban/rural areas in the Shanxi chronic disease health surveillance sites. This study was approved by the ethics committee of Shanxi Provincial Center for Disease Control and Prevention (Number: SXCDCIRBPJ2019036001). The inclusion criteria could be listed as follows: (1) type 2 diabetes diagnosis following 1999 WHO diagnostic criteria for T2DM(11); (2) signed informed consent. The exclusion criteria comprise: (1) mental disorders; (2) pregnant women; (3) patients with serious acute or chronic diseases (4) patients with respiratory infectious diseases; (5) other patients unable to complete the information collection and refused to sign the informed consent form. All surveys were completed by 1830 patients, excluding 85 patients who had already received DR treatment, leaving 1745 patients for final inclusion in the analysis.

2.2 Data collection

This is a cross-sectional study, and the data was obtained from a four-part survey consisting of a questionnaire, physical examination, DR examination and laboratory tests. Detailed information about the variables could be available in the supplementary table 1.

The questionnaires were prepared by the National Task Force on Chronic Disease Surveillance and were administered centrally to patients using face-to-face questioning by uniformly trained and qualified surveyors in a separate set-up questionnaire questioning area, which could not be completed by the survey respondents themselves. The survey was investigated from 21 June 2019 to 1 August 2019. The questionnaire includes basic information, lifestyle habits and diabetes history of the survey respondents. The demographic information comprises age, sex, region, educational background and marital status (cohabiting and separated).

Lifestyle habits consist of smoking, daily diet, physical activity and sleep disorders. The daily diet included alcohol consumption and consumption of various types of food, including 12 food groups such as cereals (rice and products, refined cereals such as steamed bread and noodles), mixed cereals (coarse grains such as corn, buckwheat and millet), potatoes, legumes and their products, fresh vegetables, fresh fruits, dairy and its products, animal meat, poultry, aquatic products, eggs and nuts. Using the food frequency method, patients were asked about the frequency of intake and the average intake of each of the above food groups per time in the past year, and the weekly food intake was calculated (weekly food intake equals weekly food intake frequency multiplied by average intake of each food per time). The intake of each food group was graded into low intake, normal intake and high intake as per the dietary recommendation guidelines(12). Physical activity included work, agricultural and domestic physical activity, transport physical activity, recreational and exercise activity. We also collected the occurrence of sleep disorders. The details of the questionnaire enquiry on lifestyle habits are shown in the Supplementary material.

History of DM includes the type of DM, the duration of the diabetes and history of DR. Diabetes must be diagnosed by a hospital at or above the county/district level and meet 1999 WHO diagnostic criteria for T2DM(11); as diabetes is a chronic disease and only type 1 diabetes has an acute onset, T2DM has an insidious onset and patients are usually unconscious in the early stages, so the duration of the disease is usually counted from the date of diagnosis.

The physical examination is carried out in a separate, quiet and comfortable room by a uniformly trained and qualified surveyor using a specified type of measuring instrument. The measurement is carried out in the early morning, avoiding strenuous exercise, ensuring an empty stomach, emptying the bladder and avoiding contact with sources of irritation before the measurement. Measurements include height, weight, blood pressure and pulse rate. Two investigators are assigned to each measurement. Body Mass Index (BMI) is defined as a person’s weight in kilograms divided by the square of height in meters. The subject was asked to relax and sit still for 5 minutes before the blood pressure was measured, and the subject's left arm was measured three times, one minute apart. Blood pressure was measured in millimetres of mercury (mmHg), the instrument was accurate to 1 mmHg and the results were accurate to 1 mmHg. Pulse was measured in beat per minute (bpm), the instrument was accurate to 1 bpm and the results were accurate to 1 bpm.

The DR examination is carried out by a uniformly trained and qualified professional photographer in a separate darkroom using a non-dilated fundus camera (Canon CR-2) and accompanying computer equipment to take and record fundus photographs of the patient. 2 photographs are taken of each eye: one of the focal point on the optic papilla and one of the central recess. 4 fundus photographs of each patient are collected centrally and sent to the testing centre for diagnosis by two independent ophthalmologists. This eye is scored for DR according to the 2017 Guidelines for Imaging and Reading Diabetic Retinopathy Screening in China(13), and the diagnosis is subject to a consistency test, with divergent samples being re-evaluated by the director of the testing centre as the final result. Each eye is divided into no DR and any DR based on the DR score above. as long as one eye is diagnosed with DR, the patient is judged to be a DR patient.

Fasting plasma glucose (FPG), glycosylated haemoglobin (HbA1c), total cholesterol (TC), triglyceride (TG), high-density lipoprotein (HDL), low-density lipoprotein (LDL), alanine aminotransferase (ALT), aspartate transaminase (AST), γ-glutamyl transpeptidase (γ-GGT), total protein (Tpro), Albumin (ALB), Globulin (GLB), Urea nitrogen (BUN), blood creatinine (Scr), blood uric acid (UA), urine albumin (UAlb), urine creatinine (Ucr) were collected from the patient’s fasting blood and urine samples. A/G ratio was calculated as ALB divided by GLB; UACR equals UAlb divided by Ucr.

Laboratories at each survey site use uniform testing equipment and standards, and analytical tests can only be performed after laboratory performance has been verified. The above indicators were graded and assigned values, as detailed in Table 1.

( Table 1)

2.3 Algorithms and models

2.3.1 Elastic Net

Lasso regression does not consider correlations between characteristics and is not suitable for multicollinearity variables(3); Ridge regression does not have predictors with real coefficients of zero for model selection. in 2005, Zoo and Hastie proposed the Elastic Net penalty model, which is a convex combination(14) of Lasso and Ridge regression, whose estimates can be expressed as:

The penalty function could be listed as:

The penalty is expressed as：. When β=0, the elastic net is equivalent to Lasso, while when β=1, the elastic net is equivalent to Ridge regression. The elastic net regression combines the two to give the ideal sparse model when cross-validating feature selection and to compensate for the effects of correlation between observed variables.

2.3.2 Bayesian Networks (BNs)

BNs is a probabilistic graphical model that represents a set of variables and their conditional dependencies via a DAG(4). BNs are an ideal tool for making predictions about events that occur, predicting the likelihood that any one of several possible known causes is a contributing factor. For example, a BNs can represent the probabilistic relationship between a disease and a symptom. BNs use graphical structures and network parameters to uniquely determine the joint probability distribution of the random variable = x｛x₁, x_n｝, which can be listed as:

2.3.3 MMHC algorithms

Structure learning for MMHC BNs is mainly implemented by the CB algorithm and the SS algorithm(9). The CB algorithm uses conditional independence tests to learn conditional independence constraints from the data. These constraints are in turn used to learn the structure of BNs. However, the premise is that conditional independence implies graphical separation (so that two independent variables cannot be connected by an arc). The SS-based algorithm is a generic optimization algorithm that ranks the fit scores of the network structure. the MMHC algorithm represents a hybrid algorithm that combines the CB algorithm and the SS-based algorithm(15) while using conditional independence tests as a way to find the global optimal solution in a reduced search space and thus construct the best network structure.

2.4 Statistical method

Statistical descriptive analysis of DR-related influences was performed using SPSS 26, with categorical variables expressed as rates (%). BNs structure was learned using the mmhc() function in the package “bnlearn” in R software (3.5.0). Parameter learning was performed using the maximum likelihood estimation method, implemented by Netica software (5.18).

3.1 General information

1745 individuals were finally enrolled in this study, of which 865 cases (49.6%) were male and 880 cases (50.4%) were female, aged 18-75 years, with a mean age of 56.73 years. 267 patients with T2DM complicated by DR, with a secondary rate of 15.3%. Compared with patients without DR, patients with low education, separation, potatoes, fresh fruits, poultry, nut intake, long course of diabetes, high FBG, high HbA1c, regular use of insulin, hypertension, cerebral hemorrhage, history of urinary tract infections, serum creatinine, BUN, UA, UACR higher than normal, and decreased A/G ratio of secondary DR were higher, as shown in Table 1. Detailed information about other variables could be available in supplementary Table 1.

3.2 Feature selection using Elastic Net

The 72 variables collected that could be associated with concurrent DR in T2DM were included in the Elastic Net model, and the optimal model parameter values were selected using a ten-fold cross-validation method to determine the penalty coefficient (λ = 0.04255 and α = 0.5) The lowest impact eigencoefficient was reduced to 0 and removed from the model, resulting in 11 variables identified as strong correlates of concurrent DR in patients with T2DM. The selected variables and their regression coefficients are shown in Supplementary Table 2 and Figure 1.

3.3 Bayesian Networks

The BNs model for factors associated with DR in patients with T2DM was further constructed using the MMHC algorithm using the 11 factors screened by Elastic Net. The constructed model is shown in Figure 2, which comprises 12 nodes and 15 directed edges. BNs suggested that the prior probability of DR stands at 0.155. Besides, BNs showed that SBP, FPG, HbA1c, BUN and UACR levels are directly related to DR, namely, SBP, FPG and HbA1c are the parent nodes of DR, and BUN and UACR are the children of DR. When the above parameters leaves at abnormal, the probability of DR reaches at 0.741. Marital status and poultry meat intake were indirectly related to DR through SBP, insulin use was directly related to diabetes duration and HbA1c, and diabetes duration was indirectly related to DR through HbA1c.

3.4 Bayesian reasoning

The conditional probability distribution table of the DR node was shown in Supplementary Table 3, which shows the probability dependence between DR and its four parent nodes. When a patient with T2DM has normal HbA1c, FPG and SBP, his risk of concurrent DR is 0.05222; when the patient presents with only high SBP symptoms and normal HbA1c and FPG, his probability of concurrent DR increases to 0.07595; when the patient presents with HbA1c, FPG and SBP all above the critical value, his risk of concurrent DR increased to 0.29032.

BNs can also make inferences about the probability of unknown nodes from known node information. We can determine the risk of concurrent DR in patients with T2DM from the order in which the physicians obtained the information. The prior probability from BNs predicted that the patient's risk of developing DR was 0.155 (Figure 2). After basic information entry and history taking, it was found that the patient was divorced, the patient had been diabetic for more than 10 years and had been using insulin regularly, and the risk of complicating DR for him was 0.214 as indicated by BNs analysis (Figure 3); after further clinical and biochemical examinations, it was learned that the The patient's risk of complicating DR was 0.488 when he had hypertension and a higher than normal BUN, but normal A/G, heart rate and FPG; if the patient also had massive proteinuria and a higher than normal HbA1c, his likelihood of complicating DR was 0.747 (Figure 4).

DR represents one of the major microvascular complications of DM and is a fundus lesion with specificity. With growing prevalence of DM globally, DR the most important cause of blindness in working-age adults poses a serious challenge to disease prevention and control in China. The incidence of DR among outpatients and inpatients with T2DM in Yunnan was 23.6%(16) in 2003, and it has been shown that the incidence of DR in China stood at 25%(17). This study showed that the secondary prevalence of DR was 15.3%. The reasons could be listed as follows: (1) the individuals in this study concern the population undergoing regular health surveillance at chronic disease survey sites in China, where the underdiagnosis rate of DM is lower and more reflective of the true prevalence; (2) the population are subjected to community surveillance, where patients have better compliance and are relatively mild and in good physical condition. The majority of the studies were based on medical records from hospital outpatient or inpatient cases with relatively poor medical and physical conditions, which contributed to the discrepancy in results. A population-based survey reported a prevalence of 11.9%(18) for DR in DM, possibly because the population was selected to avoid bias in hospital admission rates, which ultimately led to the difference in results; (3) long duration of disease is a risk factor for DR in a large number of studies, and the patients included in this study may have had a shorter duration of DM than other studies, resulting in a lower prevalence rate; (4) different prevalence rates reporting may also be due to different types of study designs and sample sizes.

In this study, a complex network linkage between DR concurrent with T2DM and various related factors were identified by constructing a model of BNs based on the MMHC algorithm. Our study suggested SBP, FPG, HbA1c, BUN and UACR levels were directly associated with DR, which is consistent with previous studies.

In previous epidemiological studies, higher FPG, higher HbA1c and higher SBP have been identified(16, 19, 20) as risk factors for DR complicating T2DM. FPG is the patient's immediate fasting glucose level at the time of blood draw and HbA1c reflects the patient's recent glycemic control over a 2-3 month period(21), they are different indicators reflecting the patient's blood glucose level. In an early clinical trial(22), it was demonstrated that patients with lower blood glucose and especially lower HbA1c levels were at lower risk of developing DR than those with high blood glucose and higher HbA1c levels. Hyperglycaemia contributes to DR mainly through mechanisms such as altering redox status(23), producing inflammatory cytokines(24), inducing fibrosis formation(25, 26) and causing retinal neurodegeneration(27), and research into the mechanisms involved is still ongoing. The retinal capillary hyperperfusion pressure produced by high SBP can damage retinal capillary endothelial cells, thereby triggering DR. Of the factors included in this paper, poultry intake and marital status were indirectly associated with DR through hypertension. It has been suggested that poultry intake may be associated with primary hypertension(28), i.e. high poultry intake is associated with reduced primary hypertension, which may be related to its high-quality protein content, low-fat content and high vitamin content. Previous studies have found that marriage is negatively associated with mean arterial pressure(29), i.e. married, cohabiting marital status is associated with decreased blood pressure. In this paper, BNs model found that patients with cohabiting marital status had a lower risk of secondary DR than patients who were unmarried, widowed or separated, suggesting that good marital status reduces the probability of DR during its development.

Previous studies have found that an increase in BUN and a decrease in UACR are associated with(30-32) increased risk of developing diabetes. BUN is an end product of protein metabolism and its excretion from the body is dependent on normal glomerular function. The kidney is an important organ for maintaining glucose homeostasis in the body. When chronic kidney disease occurs in the body, BUN accumulates in the body, and the increase in BUN not only reduces insulin secretion(33) from the β cells, but also produces a large amount of ROS by inducing oxidative stress, which causes insulin resistance(34) in the body and promotes the development of T2DM. ROS can also be induced by hyperglycemic states in the body and cause endothelial dysfunction, which in turn leads to the development of DR (23). UACR can be used to reflect renal endothelial cell function, and when patients develop microproteinuria, i.e. UACR at 30-300mg/L is considered an early indicator(35) of the onset of diabetic nephropathy, and the hyperglycaemic state can cause vascular endothelial cell damage(36) in different parts of the body, which may serve as the reason why previous studies showed UACR is associated with DR. A/G is usually used to reflect the synthetic function of the liver, and a reduced A/G ratio is usually indicative of impaired liver function. It has been shown that high glucose induces fibrosis in organs by stimulating different fibrotic pathways(26), especially in the liver and kidneys, thus causing changes in liver and kidney function. However, the results of the questionnaire in this study did show that there was no statistical association between concomitant liver and kidney disease and the occurrence of DR (P < 0.05), suggesting that BUN, UACR and A/G in the BNs constructed in this paper could provide a reference for the early diagnosis of DR and liver and kidney injury in patients with DM.

The BNs model shows that insulin use is associated with the duration of diabetes and HbA1c, and that the duration of diabetes is indirectly associated with DR through HbA1c. In patients with T2DM who suffer the disease for a long time, have severe damage to pancreatic beta cells and have poor glycaemic control with oral hypoglycaemic agents, the guidelines recommend insulin for treatment(37). In turn, prolonged duration of diabetes and poor glycaemic control is associated with the development of DR. Perhaps it could help explain why some studies(19, 38) have found an association between insulin use and the development of DR. Obviously, BNs can truly reveal the influence and interrelationship of various factors on DR through an intuitive visual approach.

Of note, in the process of disease development, various factors related to disease occurrence and clinical indicators do not exist singularly and independently of each other, but there are often complex and interrelated network relationships. In this paper, we used Elastic Net to screen the factors closely related to T2DM concurrent DR and then constructed a BNs model, which better presented the complex network connections between variables and revealed the probabilistic dependence between T2DM concurrent DR and its direct and indirect risk factors. Furthermore, as BNs can infer the probability of occurrence of events at unknown nodes when only a few nodes are known, the presentation of the probability of disease occurrence can intuitively provide patients and clinicians with an understanding of the factors influencing disease occurrence, making it more favoured by clinicians than logistic models for predicting disease occurrence in clinical work, enabling diabetic patients and clinicians to make a diagnosis. Although the directed edges represents a probabilistic dependency(39), not a causal relationship, between two nodes, which are to achieve a global optimal solution, it offers the exploration of pathophysiological mechanisms of disease occurrence and provides ideas for clinical treatment and disease prevention.

This study should be interpreted in the context of several limitations. Firstly, the patients included in this study were collected from some regions of Shanxi Province, which may be subjected to selection bias, and the conclusions drawn are not representative of the population elsewhere. Secondly, this study used a questionnaire to collect information from patients, which require respondents to recall, are prone to recall bias. In the next study, we will further include more samples from other regions to perfect our findings.

By establishing a BNs with MMHC, this study was able to achieve probabilistic inference of risk factors and risk reasoning, revealing complex network connections and probability-dependent relationships among individual risk factors, providing clues for the exploration of DR occurrence mechanisms and providing a scientific and effective basis for the prevention of DR.

Diabetic retinopathy (DR)

Bayesian networks (BNs)

type 2 diabetes mellitus (T2DM)

odds ratio (OR)

direct acyclic graphs (DAG)

conditional probability tables (CPT)

Constraint based (CB)

Scoring and searching (SS)

Max-Min Hill Climbing (MMHC)

Body Mass Index (BMI)

millimetres of mercury (mmHg)

beat per minute(bpm)

Fasting plasma glucose (FPG)

glycosylated haemoglobin (HbA1c)

total cholesterol (TC)

triglyceride (TG)

high-density lipoprotein (HDL)

low-density lipoprotein (LDL)

alanine aminotransferase (ALT)

aspartate transaminase (AST)

γ-glutamyl transpeptidase (γ-GGT)

total protein (Tpro)

Albumin (ALB),

Globulin (GLB)

Urea nitrogen (BUN)

blood creatinine (Scr)

blood uric acid (UA)

urine albumin (UAlb)

urine creatinine (Ucr)

ALB divided by GLB (A/G);

UAlb divided by Ucr. (UACR)

Ethics approval and consent to participate

This study was approved by the ethics committee of Shanxi Provincial Center for Disease Control and Prevention (Number: SXCDCIRBPJ2019036001). All the methods of this study was conducted according to the guidelines laid down in the Declaration of Helsinki and all procedures involving research study participants were approved by the ethics committee of Shanxi Provincial Center for Disease Control and Prevention. Written informed consent was obtained from all the subjects.

Consent for publication

Not applicable

Availability of data and materials

The datasets used and/or analysed during the current study are available from the corresponding author Jing Chen on reasonable request.

Competing interests

The authors declare that they have no competing interests

Funding

This work was supported by the "Four batches of innovation plan for medical revitalization with Science and Technology "

Authors' contributions

Jian Gao drafted the manuscript, Ruimin Zhang, Wenzhu Song, Yongliang Feng and Peigang Zhang helped polish the manuscript, Wenzhu Song gave precious advice on the statistical methods, and Jing Chen was responsible for the conception and design of the research.

Acknowledgements

We show our sincere gratitude for those who help us a lot during our manuscript preparation. We also indebeted to the funding.

Wang Y, Zhao X, Jiang Y, Li H, Wang L, Johnston SC, et al. Prevalence, knowledge, and treatment of transient ischemic attacks in China. Neurology. 2015;84(23):2354-61.
Pan J, Rao H, Zhang X, Li W, Wei Z, Zhang Z, et al. Application of a Tabu search-based Bayesian network in identifying factors related to hypertension. Medicine (Baltimore). 2019;98(25):e16058.
George EI. The Variable Selection Problem. J Am Stat Assoc. 2000;95(452):1304-8.
Chadegani AA, Poursina D. An examination of the effect of discretization on a nave Bayes models performance. Academic Journals. 2013.
Mancini F, Sousa FS, Hummel AD, Falcao AE, Yi LC, Ortolani CF, et al. Classification of postural profiles among mouth-breathing children by learning vector quantization. Methods Inf Med. 2011;50(4):349-57.
C L. Independency relationships and learning algorithms for singly connected networks. Journal of Experimental and Theoretical Artificial Intelligence. 1998;4:511-49.
D H. Learning bayesian networks: The combination of knowledge and statistical data. Machine Learning. 1995;20:197-243.
Cooper GF. A Bayesian Method for the induction of probabilistic networks from data.pdf>. Machine Learning. 1992: 309-47.
Tsamardinos I, Brown LE, Aliferis CF. The max-min hill-climbing Bayesian network structure learning algorithm. Machine Learning. 2006;65(1):31-78.
Zou H, Hastie T. Regularization and Variable Selection via the Elastic Net. Journal of the Royal Statistical Society Series B (Statistical Methodology). 2005;67(2).
Alberti KG, Zimmet PZ. Definition, diagnosis and classification of diabetes mellitus and its complications. Part 1: diagnosis and classification of diabetes mellitus provisional report of a WHO consultation. Diabet Med. 1998;15(7):539-53.
Yue-xin Y, Huan-mei Z. Introduction to dietary guidelines for Chinese residents (2016). Acta Nutrimenta Sinica. 2016;38(03):209-17.
Society CO, Association OboCM. Guidelines for Imaging and Reading Diabetic Retinopathy Screening in China（2017）. Chinese Journal of Ophthalmology. 2017;53(12):890-6.
Garcia-Carretero R, Vigil-Medina L, Barquero-Perez O, Mora-Jimenez I, Soguero-Ruiz C, Goya-Esteban R, et al. Logistic LASSO and Elastic Net to Characterize Vitamin D Deficiency in a Hypertensive Obese Population. Metab Syndr Relat Disord. 2020;18(2):79-85.
Wang X, Pan J, Ren Z, Zhai M, Zhang Z, Ren H, et al. Application of a novel hybrid algorithm of Bayesian network in the study of hyperlipidemia related factors: a cross-sectional study. BMC Public Health. 2021;21(1):1375.
Lei L, uan-ming XY, Ya-ping K, Yu S, Yun Z, Shan M. Prevalence of diabetic retinopathy and its correlated risk factors of 1254 type 2 diabetic patients. Chin J Diabetes. 2005(05):338-9.
Yang QH, Zhang Y, Zhang XM, Li XR. Prevalence of diabetic retinopathy, proliferative diabetic retinopathy and non-proliferative diabetic retinopathy in Asian T2DM patients: a systematic review and Meta-analysis. Int J Ophthalmol. 2019;12(2):302-11.
Hu Y, Teng W, Liu L, Chen K, Liu L, Hua R, et al. Prevalence and risk factors of diabetes and diabetic retinopathy in Liaoning province, China: a population-based cross-sectional study. PLoS One. 2015;10(3):e0121477.
Ying C, Hai-Ke G, Yun-Fei H, Qian-LI M, Liang Z, Dong-Ming Y. Prevalence and risk factors of diabetic retinopathy in type 2 diabetic inpatients. Rec Adv Ophthalmol. 2012;32(08):736-9.
Xiang G. Relevant risk factors of diabetic retinopathy. Chin J Exp Ophthalmol. 2003(03):299-301.
American Diabetes A. 6. Glycemic Targets: Standards of Medical Care in Diabetes-2021. Diabetes Care. 2021;44(Suppl 1):S73-S84.
The absence of a glycemic threshold for the development of long-term complications: the perspective of the Diabetes Control and Complications Trial. Diabetes. 1996;45(10):1289-98.
Madsen-Bouterse SA, Zhong Q, Mohammad G, Ho YS, Kowluru RA. Oxidative damage of mitochondrial DNA in diabetes and its protection by manganese superoxide dismutase. Free Radic Res. 2010;44(3):313-21.
King GL. The role of inflammatory cytokines in diabetes and its complications. J Periodontol. 2008;79(8 Suppl):1527-34.
Lomonaco R, Godinez Leiva E, Bril F, Shrestha S, Mansour L, Budd J, et al. Advanced Liver Fibrosis Is Common in Patients With Type 2 Diabetes Followed in the Outpatient Setting: The Need for Systematic Screening. Diabetes Care. 2021;44(2):399-406.
Tuleta I, Frangogiannis NG. Diabetic fibrosis. Biochim Biophys Acta Mol Basis Dis. 2021;1867(4):166044.
Frank T, Nawroth P, Kuner R. Structure-function relationships in peripheral nerve contributions to diabetic peripheral neuropathy. Pain. 2019;160 Suppl 1:S29-S36.
Li C, Jihong H, Limei J, Jia Y, Tingting M, Shunan Z, et al. Study on the relationship between dietary habits and essential hypertension of Tibetan people in Gannan Tibetan Autonomous Prefecture. J GanSu University of Chinese Medicine. 2018;35(01):118-21.
Rong-sheng L, Ding-yu Y, Xiao-hui L, Zu-qing X, Hong-wen Z, Ming T, et al. Comparative epidemiological study on blood pressure and psychosocial factors in urban and rural population. Chin J of Behavioral Med Sci. 1995(02):66-8+111.
Li-Tao G, Li-Min L, -Yuan ZY, Lei L, -Dong HYe, Lei C. Risk factors analysis of diabetic retinopathy. Rec Adv Ophthalmol. 2011;31(08):742-4.
Qing-Fang Z, Hui-Yu W, Li-Ying H. Correlation between the degree of diabetic retinopathy and renal function in patients with type 2 diabetes mellitus. Rec Adv Ophthalmol. 2018;38(05):471-4.
Xie Y, Bowe B, Li T, Xian H, Yan Y, Al-Aly Z. Higher blood urea nitrogen is associated with increased risk of incident diabetes mellitus. Kidney Int. 2018;93(3):741-52.
Koppe L, Nyam E, Vivot K, Manning Fox JE, Dai XQ, Nguyen BN, et al. Urea impairs beta cell glycolysis and insulin secretion in chronic kidney disease. J Clin Invest. 2016;126(9):3598-612.
D'Apolito M, Du X, Zong H, Catucci A, Maiuri L, Trivisano T, et al. Urea-induced ROS generation causes insulin resistance in mice with chronic renal failure. J Clin Invest. 2010;120(1):203-13.
Caramori ML, Fioretto P, Mauer M. The need for early predictors of diabetic nephropathy risk: is albumin excretion rate sufficient? Diabetes. 2000;49(9):1399-408.
Ahsan H. Diabetic retinopathy--biomolecules and multiple pathophysiology. Diabetes Metab Syndr. 2015;9(1):51-4.
Society CD. Guidelines for prevention and treatment of type 2 diabetes in China (2013 Edition). Chin J Diabetes. 2015.
Na L, Xiu-fen Y, Yu D, Hong G, Xue-tao R, Jun X, et al. Diabetes self-management and its association with diabetes retinopathy in patients with type 2 diabetes. China J Ophthalmol. 2013(06):500-6.
Stepanenko S, Engels B. Tabu search based strategies for conformational search. J Phys Chem A. 2009;113(43):11699-705.

Table 1 Baseline characteristics

Variables	Without DR (N=1478)	With DR (N=267)
Age
＜35years	51 (3.5%)	4 (1.5%)
35-54 years	453 (30.6%)	74 (27.7%)
＞54 years	974 (65.9%)	189 (70.8%)
Sex
Male	728 (49.3%)	137 (51.3%)
Female	750 (50.7%)	130 (48.7%)
Region
Xinghualing（urban）	390 (26.4%)	55 (20.6%)
Huguan（rural）	353 (23.9%)	71 (26.6%)
Yuci（urban）	391 (26.5%)	83 (31.1%)
Jiangxian（rural）	344 (23.3%)	58 (21.7%)
Marital status
divorced and widowed	108 (7.3%)	34 (12.7%)
cohabitation	1370 (92.7%)	233 (87.3%)
Income
＜5000 Yuan/month	328 (22.2%)	69 (25.8%)
≥5000 Yuan/month	363 (24.6%)	62 (23.2%)
≥10000 Yuan/month	396 (26.8%)	73 (27.3%)
≥20000 Yuan/month	391 (26.5%)	63 (23.6%)
Smoke
YES	558 (37.8%)	103 (38.6%)
NO	920 (62.2%)	164 (61.4%)
Drink
YES	469 (31.7%)	79 (29.6%)
NO	1009 (68.3%)	188 (70.4%)
Duration of diabetes
Less than 5 years	682 (46.1%)	89 (33.3%)
More than 5 years	485 (32.8%)	92 (34.5%)
More than 10 years	311 (21%)	86 (32.2%)
Use insulin regularly
Yes	216 (14.6%)	73 (27.3%)
No	1262 (85.4%)	194 (72.7%)
BMI
＜18.5 kg/m2	13 (0.9%)	4 (1.5%)
≥18.5 kg/m2	438 (29.6%)	88 (33%)
≥24 kg/m2	658 (44.5%)	124 (46.4%)
≥28 kg/m2	369 (25%)	51 (19.1%)
Waist
Male＜85cm， Female＜80cm	336 (22.7%)	62 (23.2%)
Male≥85cm， Female≥80cm	1142 (77.3%)	205 (76.8%)
SBP
＜140mmHg	1039 (70.3%)	152 (56.9%)
≥140mmHg	439 (29.7%)	115 (43.1%)
DBP
＜90mmHg	1250 (84.6%)	224 (83.9%)
≥90mmHg	228 (15.4%)	43 (16.1%)
Heart rate
＜60bpm	66 (4.5%)	5 (1.9%)
60-90bpm	1351 (91.4%)	240 (89.9%)
＞100bpm	61 (4.1%)	22 (8.2%)
FPG
≤7.0mmol/L	587 (39.7%)	47 (17.6%)
＞7.0mmol/L	891 (60.3%)	220 (82.4%)
HbA1c
＜7.0%	743 (50.3%)	61 (22.8%)
≥7.0%	735 (49.7%)	206 (77.2%)
TC
＜4.5mmol/L	696 (47.1%)	114 (42.7%)
≥4.5mmol/L	782 (52.9%)	153 (57.3%)
TG
＜1.7mmol/L	689 (46.6%)	123 (46.1%)
≥1.7mmol/L	789 (53.4%)	144 (53.9%)
HDL
Male＜1.0mmol/L， Female＜1.3mmol/L	772 (52.2%)	126 (47.2%)
Male≥1.0mmol/L， Female≥1.3mmol/L	706 (47.8%)	141 (52.8%)
LDL
＜2.6mmol/L	663 (44.9%)	108 (40.4%)
≥2.6mmol/L	815 (55.1%)	159 (59.6%)
ALT
Male≤41U/L， Female≤31U/L	747 (50.5%)	143 (53.6%)
Male＞41U/L， Female＞31U/L	731 (49.5%)	124 (46.4%)
AST
Male≤38U/L， Female≤32U/L	788 (53.3%)	147 (55.1%)
Male＞38U/L， Female＞32U/L	690 (46.7%)	120 (44.9%)
γ-GGT
Male≤60U/L， Female≤40U/L	749 (50.7%)	134 (50.2%)
Male＞60U/L， Female＞40U/L	729 (49.3%)	133 (49.8%)
Tpro
＜65g/L	3 (0.2%)	2 (0.7%)
65-85g/L	1337 (90.5%)	233 (87.3%)
＞85g/L	138 (9.3%)	32 (12%)
Alb
＜40g/L	4 (0.3%)	6 (2.2%)
40-55g/L	1426 (96.5%)	253 (94.8%)
Glob
＞55g/L	48 (3.2%)	8 (3%)
＜20g/L	5 (0.3%)	2 (0.7%)
20-40g/L	1456 (98.5%)	264 (98.9%)
＞40g/L	17 (1.2%)	1 (0.4%)
TG
＜1.7mmol/L	689 (46.6%)	123 (46.1%)
≥1.7mmol/L	789 (53.4%)	144 (53.9%)
BUN
≤8.3mmol/L	1450 (98.1%)	245 (91.8%)
＞8.3mmol/L	28 (1.9%)	22 (8.2%)
Scr
Male＜59μmol/L， Female＜45μmol/L	58 (3.9%)	13 (4.9%)
Male≥59μmol/L， Female≥45μmol/L	1354 (91.6%)	226 (84.6%)
Male＞104μmol/L，Female＞84μmol/L	66 (4.5%)	28 (10.5%)
UA
Male＜202μmol/L，Female＜142μmol/L	42 (2.8%)	16 (6%)
Male≥202μmol/L，Female≥142μmol/L	1107 (74.9%)	201 (75.3%)
Male＞416μmol/L，Female＞339μmol/L	329 (22.3%)	50 (18.7%)
A/G
＜1.5	264 (17.9%)	81 (30.3%)
≥1.5	1214 (82.1%)	186 (69.7%)
UACR
＜30mg/g	387 (26.2%)	33 (12.4%)
30-300mg/g	834 (56.4%)	131 (49.1%)
＞300mg/g	257 (17.4%)	103 (38.6%)

#Table 1 should appear at line 162

No competing interests reported.

Supplementarymaterial.docx
Name: Supplementary material File format: .docx Title of data: Supplementary material Description of data: Including the completed Table1 and detailed definitions related to the questionnaire.

Bayesian network-based risk factors analysis for diabetic retinopathy: A cross-sectional study

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methods

2.1 Subjects

2.2 Data collection

2.3 Algorithms and models

2.4 Statistical method

3. Results

3.1 General information

3.2 Feature selection using Elastic Net

3.3 Bayesian Networks

3.4 Bayesian reasoning

4. Discussion

5. Conclusion

6. List of Abbreviations

Declarations

References

Tables

Additional Declarations

Supplementary Files

Status:

Version 1