Establishment of a Risk Prediction Model for Non-alcoholic Fatty Liver Disease in Type 2 Diabetes

The study is a retrospective cohort study. Based on screening of the electronic medical records system of the Sanlin and Jinyang community health service centers in Shanghai, we identified 24 relevant indicators of T2DM. In 2016–2017, data on 1013 people with T2DM were collected, and these data were screened and followed up in these two communities from September 2018 to August 2019. Of the 1013 patients with T2DM originally identified, 139 were not included in the final analysis due to loss to follow-up, missing or incorrect data and/or missing samples (loss rate 13.7%). The final patient population for assessment thus included 874 people with T2DM. A flow diagram of the study design is shown in Fig. 1.

The patients with T2DM eligible for entry into the study had been admitted to the Shanghai University of Traditional Chinese Medicine, School of Public Health between September 2018 and August 2019. Inclusion criteria were: Exclusion criteria were: (1) presence of cognitive impairment and/or inability to participate in the cohort independently; (2) presence of serious organic disease, such as a tumor, major surgery, etc.

All analyses were approved by the School of Public Health, Shanghai University of Traditional Chinese Medicine and Sanlin/Jinyang Community Health Centers. All procedures were carried out following international guidelines and regulations by trained researchers. All researchers are doctors, nurses or researchers with professional medical training.

The testing process consisted of:

Patients’ demographic characteristics, physical data, serum biomarkers and transabdominal ultrasound data were collected, and all data were then submitted electronically to the Diabetes Data Management System of our group. All patients completed a questionnaire survey to provide basic demographic variables including age, sex, course of T2DM, drugs, smoking status, alcohol drinking status, among others. The questionnaire was given as an electronic questionnaire on a tablet, with the staff entering the questionnaire into the questionnaire system in advance and helping each patient at one-on-one to obtain the questionnaire data.

A physical examination was also conducted to collect data on body weight (kg), height (m), waistline (cm), systolic blood pressure (SBP) and diastolic blood pressure (DBP) (SBP and DBP measurements [in mmHg] were obtained using an automatic electronic blood pressure monitor [model HBP-9021; Omron Corp., Kyoto, Japan]). BMI was calculated as weight (kg)/(height [m])². An abdominal ultrasound examination was carried by a specialist on the iU22 xMATRIX Ultrasound system (Philips N.V., Amsterdam, the Netherlands). Four technicians were involved in the ultrasound examinations, all of whom were trained to diagnose NAFLD using the same criteria used prior to the study to ensure quality control of the diagnostic images and results. One physician and one nurse also participated in the scan.

Blood samples were collected to obtain essential biochemical variables, including fasting blood glucose (FBP), postprandial blood glucose (PBG), hemoglobin A1c (HbA1c), total cholesterol (TC), triglyceride (TG), low-density lipoprotein-cholesterol (LDL-C), high-density lipoprotein-cholesterol (HDL-C), blood urea nitrogen (BUN), serum creatinine (SCR), serum uric acid (SUA), urinary creatinine (UCR), urine microalbumin (UMA), urine albumin and creatinine ratio (ACR). Blood samples were collected through standardized processes and stored under standard conditions (refrigeration at 4 °C) until sent to the laboratory of Ruijin Hospital in Luwan District for measurement (within 2 h of collection). An automatic biochemical analyzer was used for biochemical testing (AU5800 clinical chemistry analyzer; Beckman Coulter Inc., Brea, CA, USA).

The previous study and data collection were approved by the Ethics Committee of Shanghai Oriental Hospital affiliated to Tongji University (Batch number: [2017] Research Review No. 20). All patients gave informed consent and signed paper informed consent. The study was conducted in accordance with the 1964 Declaration of Helsinki and its subsequent amendments.

Statistical analysis was performed using R software (version 3.6.1; R Foundation for Statistical Computing, Vienna, Austria), First, the R caret package was used to randomly divide the 874 patients with T2DM into a training set with 658 participants and a validation set with 216 participants for external validation, conforming to the theoretical ratio of 3:1 [14]. The glmnet package was used to run LASSO (least absolute shrinkage and selection operator) regression analysis, which is a shrinkage and variable selection method for linear regression models. To obtain the subset of predictors, the LASSO regression analysis minimizes prediction error for a quantitative response variable by imposing a constraint on the model parameters that cause the regression coefficients for some variables to shrink toward zero. Variables with a regression coefficient equal to zero after the shrinkage process are excluded from the model while variables with non-zero regression coefficients variables are most strongly associated with the response variable. We set family = “binomial”; this parameter is suitable for the binary discrete dependent variable (binary) because the included dependent variable is whether NAFLD is present or not. Then we set type.measure = deviance to use deviance, that is − 2log-likelihood. Based on the type measure of − 2log-likelihood and the binomial family, the LASSO regression analysis running in R software runs a k-fold (tenfold in this case) cross-validation for centralization and normalization of the included variables and then picks the best lambda value. Lambda.lse gives a model with good performance but the least number of independent variables. So the LASSO method was used to analyze the data in the training set to select the optimal predictors of the present risk factors, including sex, age, course of disease, hypertension, diabetic retinopathy (DR), diabetic nephropathy (DN), BMI, waistline, SBP, DBP, FBG, PBG, HbA1c, TC, TG, LDL-C, HDL-C, BUN, SCR, UA, UCR, UMA and ACR. The risk factor variables were initially screened using these above-mentioned inclusion variables.

Then we used the R language rms package to do logistic regression; a multivariable logistic regression analysis was used to construct a predictive model by introducing the features selected in the LASSO regression model [15]. The features were considered to be odds ratios and a P value with 95% confidence interval. The statistical significance levels were all two-sided. Introducing all the selected features and analyzing statistical significance levels of the features, we applied the statistically significant predictors to develop a predictive model for NAFLD risk. In our study, all of the selected features had statistical significance and were applied to develop the nomogram prediction models. The rms package used to develop the nomogram diagram also uses the R language.

Further, several kinds of validation methods were used to estimate the accuracy of the risk prediction model by using the data of the training set and the validation set respectively. We used the R language pROC package for receiver characteristic curve (ROC) operation. The area under the ROC was used to provide good discrimination for the quality of the risk nomogram to separate true positives from false positives [16]. We used the rms package to draw and calculate the calibration curves, which were used to evaluate the calibration of the NAFLD risk nomogram, accompanied by the Hosmer–Lemeshow test (HLtest.R). We used the nricens package for the decision curve analysis (DCA), which is used to determine the clinical practicability of nomograms based on the net benefit under different threshold probability in the T2DM cohort [17].

Of the 874 participants with T2DM in our analysis, of whom 394 were men and 480 were women, 415 had NAFLD and 459 did not have NAFLD. The prevalence of NAFLD in this cohort was thus 47.5%. These patients were randomly placed into the training set or validation set at a ratio of 3:1 (3 in the training set and 1 in the validation set), such that 658 and 216 patients were included in the training set and validation set. All patients completed the related examinations. The basic characteristics of the patients in the two sets are given in Table 1.

Among the patients with NAFLD and without NAFLD, 38.1 and 51.4% were men, respectively. In patients with NAFLD, 27.5% had hypertension, 18.3% had DR and 41.2% had DN; in patients without NAFLD, 52.5% had hypertension, 16.1% had DR and 30.9% had DN.

Multivariate logistic regression analysis demonstrated that among our patients with T2DM, sex, age, course of disease, BMI, waistline, DBP, TG, HDL-C and SUA were independent risk factors for NAFLD (Fig. 2).

LASSO regression analysis was used to select predictive variables from those shown in Table 1, and the multivariate logistic regression was used to establish the predictive model. Nine of the original 24 variables were included in the predictive model, namely sex, TG, BMI, waistline, DBP, SUA, course of disease, age and HDL-C, as predictors (Fig. 2). These nine variables had nonzero coefficients in the LASSO regression model. The predictive model was presented as a nomogram, which was used to quantitatively predict the risk probability of NAFLD in patients with T2DM (Fig. 2).

The results of the logistic regression analysis of these nine variables are given in Table 2. As all nine predictors showed significant statistical differences, i.e. they were independent of each other, they were introduced into the predictive model to develop a NAFLD risk nomogram (Fig. 3a). For example, using the nomogram model, a male patient with T2DM of age 62 years, a duration of diabetes for 15 years, a BMI of 28.16 kg/m², waistline measurement of 84 cm, a DBP of 94 mmHg, TG of 1.14 mmol/L, HDL-C of 1.07 mmol/L and SUA of 269 μmol/L has an estimated probability of NAFLD of 28.4% (Fig. 3b).

The ROC curve was used to evaluate the discriminatory capacity of the predictive model. For the predictive model, the pooled area under the ROC of the nomogram is 0.848 in the training set and 0.809 in the validation set (Fig. 4), which indicates moderately good performance (Fig. 4).

A calibration plot and Hosmer–Lemeshow test were used to calibrate the predictive model. From the calibration curves, the predictive model and the validation set showed a very good degree of fit. As shown by the Hosmer–Lemeshow test, the predicted and actual probability were highly consistent (training set, P = 0.906; validation set, P = 0.049) (Fig. 5).

The DCA demonstrated that the threshold probability of the prediction model in the training set and validation set is 48–91% and 44–82%, respectively (Fig. 6).

The impact of various risk factors for complications in T2DM may differ by sex. Men with T2DM appear to suffer more microvascular complications, while women have higher morbidity and mortality associated with macroangiopathy disease and also fare worse psychologically [23]. Some studies have shown that sex distinctions do exist in T2DM patients and NAFLD patients. Arnetz et al. reported that the presence of diabetes confers the potential for a greater risk for vascular complications in women compared with men due to a number of factors, including the contribution of sex hormones and sex-specific risk factors [23]. At any given age, women have about half the risk of cardiovascular disease relative to men, and estrogens can collectively influence obesity, inflammatory disease, diabetes, cancers, metabolic syndrome and NAFLD [24]. Interestingly, our results show that men with diabetes are more likely to have NAFLD than women. More epidemiological research needs to be done to confirm this association.

High TG levels and low HDL-C levels are common in patients with NAFLD [9]. Patients with higher lipid ratios (TC:HDL-C or TG:HDL-C ratios) have been found to have a significantly greater risk for advanced NAFLD and diabetes [25]. In a large, cross-sectional study conducted among 44,767 Taiwanese patients, the overall prevalence rate of NAFLD in the patient cohort was 53.76%; however, the NAFLD prevalence rate for those with the lowest TC:HDL-C and TG:HDL-C ratios was 33.41% in comparison to the prevalence rate of 78.04% in the group with the highest ratios [9]. It has also been reported that HDL-C may directly alter glucose metabolism by promoting pancreatic β-cell insulin secretion [26], and epidemiological studies and laboratory analyses have demonstrated that low levels of HDL-C are associated with a higher risk of T2DM [27, 28].

In a meta-analysis study of global epidemiology, obesity prevalence among patients with NAFLD was estimated at 51.3% [29], and statistics in the National Diabetes Statistics Report indicate that the prevalence of obesity/overweight in diabetes is as high as 87.5% in adults with diabetes in the USA [30]. Obesity has been reported to be an independent risk factor for both T2DM [31] and NAFLD [9]. An excessive BMI and visceral obesity are the most common and well-documented risk factors for NAFLD, also contributing to metabolic syndrome and other complications [32]. In diabetes, individuals with excessive fat storage have been noted to have higher insulin levels and are said to have insulin resistance, which is described as the cause of T2DM [33]. The entire spectrum of obesity, ranging from overweight to obese and severely obese, is associated with NAFLD. In our study, the proportion of patients with an excessive BMI was significantly larger in patients with T2DM and NAFLD, compared to T2DM patients without NAFLD. Our results agree with those from studies reporting that an excessive BMI and large waistline can promote the development of diabetes and NAFLD.

Interestingly, it has been demonstrated that patients with uncontrolled blood pressure despite anti-hypertensive therapy are at increased risk for developing diabetes mellitus [34]. Also, the average blood pressure level of diabetic patients is higher than that of non-diabetic patients, and elevated blood pressure levels have also been recognized as a risk factor in diabetic patients [35,36]. A systematic review based on 100,354 participants in 40 trials indicated that, among patients with T2DM, lower blood pressure is associated with improved mortality and other clinical outcomes and that patients with a baseline blood pressure of ≥ 140 mmHg have lower risk rate [37]. NAFLD has a strong association with metabolic syndrome, and blood pressure is one of the risk factors for the occurrence of metabolic syndrome [37]. Our research also confirmed that there is a close relationship between blood pressure and NAFLD in the diabetic patient population.

The level of SUA indicates the degree of metabolism of uric acid in an individual. Persons with increased SUA are at higher future risk of metabolic syndrome, T2DM and cardiovascular diseases, independent of other known risk factors, suggesting a potential role of SUA in the pathogenesis of these diseases. As reported in previous studies, SUA has emerged as an independent predictor for severity of liver damage in NAFLD development [38]. Patients with T2DM and NAFLD have a higher possibility of developing chronic kidney disease [39]. In a study of a US population, the likelihood of NAFLD increased with increasing SUA levels among the total study population, with individuals with NAFLD having higher SUA levels than those without NAFLD and the prevalence rate of NAFLD progressively increasing in parallel to increasing SUA [40]. In prospective studies, the elevation of SUA has been shown to be independently associated with a higher incidence of NAFLD after adjustment for potential confounders [41]. However, diabetic patients with persistent hyperuricemia have an increased risk of diabetic complications, especially renal disease [42]. Consequently, SUA is a major predictor of NAFLD.

Our study also confirmed that the course of disease and age are protective factors for T2DM with NAFLD in the Chinese population. Age at the time of diabetes onset and the course of diabetes are closely related to the development of NAFLD. We speculate that the earlier the onset of diabetes, the longer the course of T2DM, the higher the likelihood of developing NAFLD. The course of T2DM also shows the same trend as age—that is, the earlier the onset of diabetes, the greater the probability of NAFLD. Our results also fill in the shortcomings of published reports in this field.

Several studies have confirmed that a nomogram is a new, non-invasive, risk prediction model that is important for screening and clinic practice [38‐43]. The nomogram is now widely used in risk prediction for various diseases, including hypertension, stroke, among others [44, 45]. Application of this new model can accurately screen related variables and indicators and identify the most suitable risk factors; this model is of great significance in the primary and secondary prevention of diabetes associated with chronic diseases [15,46].

Once T2DM occurs, patients will face the burden of lifelong illness and development-related complications [47], which may lead to a serious decline in quality of life and an increased risk of death. Diabetes with NAFLD causes even more irreversible effects on the patient's metabolic system [27,48]. The current situation in China is such that elderly diabetic patients are unable to go to the hospital for a timely regular liver pathological examination due to inconvenience, difficulty in getting medical treatment, among other factors. The system of door-to-door examination by the family doctor currently promoted in China is a godsend for these patients, and the conclusions drawn from these examinations provide family doctors with reliable methods for universal screening, as well as a more scientific reference basis for treatment. Consequently, relevant indicators, such as those reported here, are of great value in practical applications [49].

First, this study was limited by funding, and the authors were unable to collect data on liver fibrosis and cirrhosis for inclusion in the analysis; application of such data will be the focus of future research. Second, based on the existing screening results, the authors identified 24 conventional indicators. In future research, we will assess some of these indicators to enrichen the established models as these may be related to the occurrence and development of NAFLD, although current information is insufficient. Finally, if the population sample size can be further expanded to the entire T2DM population in the Pudong district or to that in all of Shanghai, the established model could be applied well.