Using the multistage sampling method, from January 2018 to January 2019, three community health service centers in Hefei that signed the medical consortium agreement with the First Affiliated Hospital of Anhui Medical University were selected to screen the population aged 30 to 80 years in each community. The exclusion criteria are as follows: previous history of T2DM, severe cardiovascular and cerebrovascular diseases, liver and kidney function impairment, infection or stress, and people in the terminal stage of various diseases.

Doctors and personnel with professional training in designated physical examination institutions had undergone physical checkups based on standard procedures, such as height, weight, AGE test, fundus examination, blood pressure, and blood and urine samples.

Gender, age, height, weight, waist circumference, hip circumference, blood pressure, and other information of patients in all participants were recorded, and the body mass index (BMI) and waist–hip ratio (WHR) were calculated. Fasting venous blood was collected, and measured fast plasma glucose (FPG), 2h plasma glucose (2hPG), glycated hemoglobin A1c (HbA1c), insulin, and C-peptide were tested by the oral glucose tolerance test (OGTT); total cholesterol (TC), triglyceride (TG), high-density lipoprotein (HDL-C), creatinine, urea, and other indicators were tested by blood serum. These risk factors are collectively referred to as health risk factors (HRFs). At the same time, the non-invasive T2DM detector (National Medical Device Registration Certificate: Anhui Injection Certificate 20152210045) was used to detect the fluorescence of the subjects’ upper arm skin. Obvious blood vessels, scars, and malformed skin areas should be avoided during measurement. Selected test sites should be wiped with alcohol to obtain the fluorescence spectrum intensity of AGEs on subjects’ skin (unit: AU). Measurements should be made for 3 times, and the average value should be taken as the final skin AGE testing results. The ethics committee approval number is PJ2019-09-05.

Hypertension was characterized via systolic blood pressure (SBP) ≥ 140 mmHg and/or diastolic blood pressure (DBP) ≥ 90 mmHg (WHO 1999 criteria) (18) or currently using medications for the lowering of blood pressure. Hyperlipidemia was characterized via serum concentrations of TC>5.69 mmol/l or TG>1.68 mmol/l or HDL>1.03 mmol/l or patients that have been treated with medications for lowering of lipids. Overweight was defined as a BMI >24 and <28 kg/m². Obesity was characterized as BMI ≥ 28 kg/m² based on Chinese standards (19). An elevated level of HbA1c was greater than 7.0% (20). Hyperuricemia was characterized via serum uric acid level ≥ 416 mmol/l (≥7.0 mg/dl) in men and ≥386 mmol/l (≥6.5 mg/dl) in women or if they were treated with allopurinol for lowering the level of uric acid (21). The age limit was 65 years (22). FBG (fast blood glucose) was ≥7.0 mmol/l (23, 24).

In order to improve the effect of HRFs results, we used three different approaches to operationalize HRFs. First, we calculate a HRFs score which added together the total number of HRFs experienced by each participant. This was then categorized as 0 HRFs, 1 HRF, 2 HRFs, and 3+ HRFs. Second, factor analysis (FA) was used to derive distinct variable centered HRFs clusters, namely groups of HRFs that tend to co-occur across participants. This approach helped us to reduce the risk of overfitting. In the analysis, each FA-derived cluster was indexed by a binary variable indicating the presence or absence of any exposure to the relevant HRFs throughout childhood. Third, latent class analysis (LCA) was performed using Mplus in order to derive HRFs clusters.

All the data were analyzed using SPSS 23.0. Statistical analyses were carried out in three main steps. Missing data on the outcome and exposure variables were estimated using multiple imputations by SPSS 23.0. Latent class analysis (LCA) was used to identify homogeneous, mutually exclusive “patterns” of 6 HRFs using MPlus 7.4. Model building begins with class enumeration, and with the addition of each subsequent latent class, goodness-of-fit indices are compared. The HRF classes were adopted based on model fit indices: Akaike information criteria (AIC), Bayesian information criteria (BIC), and sample size-adjusted BIC (aBIC), entropy, and p-value for the Lo–Mendell–Rubin test (LMRT).

The specific statistical analyses were carried out in four main steps. Step 1 entailed the descriptive statistics associated with the health risk factor by AGE accumulation. In Step 2, multilevel logistic regression was used to test the association between risk factors and AGEs. In Step 3, according to the principle of LCA, the Mplus program was used to perform LCA. In Step 4, the PROCESS program of moderation was used to perform a moderation analysis (25). For testing the moderating effect, the relationships of the three small steps had to be significant: (a) direct effect of predictor (health risk factors) on AGE accumulation; (b) direct effect of moderator (BMI) on AGE accumulation; and (c) direct interactions effect (health risk factors × BMI) on AGE accumulation. In the SPSS PROCESS, the interacting effect was calculated automatically by the program, and it also produces the proportion of the variance explained by the moderating effect of BMI (R2 increases owing to interaction). Considering the influence of covariates, we also adjusted the influence of relevant sociodemographic factors in the adjustment model. We also explored whether BMI moderated the relationship between HRFs and AGE accumulation by gender (26). According to statistical requirements, the moderating effect is considered significant when the 95% CI interval does not contain zero (27).

Table 1 presents the results of the risk factor and AGE accumulation; these results show that higher HbA1c (χ ² = 25.76), FBG (χ ² = 37.38), SBP (χ ² = 8.46), and older age (χ ² = 231.11) were more likely to have a higher prevalence of AGE accumulation; others such as HDL (χ ² = 4.29), UA (χ ² = 2.71), TC (χ ² = 6.24), TG (χ ² = 3.90), and gender (χ ² = 1.94) have no correlation with AGE accumulation.

The demographic data of all participants (gender, age, nationality), physical examination (systolic pressure, diastolic blood pressure, BMI), T2DM number and type of glycemic index (fasting plasma glucose, 2 h postprandial blood glucose, glycosylated hemoglobin), renal function, blood urea nitrogen, serum creatinine, blood uric acid, and blood lipid (total cholesterol, triglycerides, high-density lipoprotein cholesterol (HDL-c)) are shown in Table 2 ; according to the multivariate variable analysis, the AGE accumulation was divided into ≤P25, P25~P50, P50~P75, and >P75 according to the quartile method. SBP, FBG, HbA1c, and age were positively correlated with AGE (OR = 0.621, 95% CI: 0.378, 0.953), (OR = 0.239, 95% CI: 0.121, 0.47), (OR = 0.243, 95% CI: 0.116, 0.507), and (OR = 0.021, 95% CI: 0.01, 0.042), respectively; TG, TC, HDL, and UA were not positively correlated with AGE (OR = 0.239, 95% CI: 0.121, 0.47), (OR = 0.239, 95% CI: 0.121, 0.47), (OR = 0.787, 95% CI: 0.357, 1.736), (OR = 0.239, 95% CI: 0.121, 0.47), and gender (OR = 1.042, 95% CI: 0.706, 1.539). These results are shown in Table 2 .

Table 3 depicts the fit statistics for the one to four class models. The five-class model did not replicate the best log likelihood value and was therefore not considered further. The four-class solution was chosen as the final, best-fitting model based on the lower AIC, as well as lowest BIC and aBIC values. Moreover, the p value of LMR was not statistically significant in class 4. Moreover, we listed the trend of aBIC values and found that class three has the most significantly descending aBIC values. With regard to model 3, bootstrap validation procedures also demonstrated that it had a good fit (p < 0.01). All remaining results are reported specific to the three-class solution.

As shown in Table 4 , all risk factor patterns were statistically significant with AGE accumulation of participants. AGE was taken as the dependent variable, with medium risk (OR: 1.50, 95% CI = 1.02, 2.32) and high risk (OR: 5.41, 95% CI = 2.42, 12.07). The results of adjusting co-variables are still meaningful, as shown in Table 4 .

In the model adjusted for gender, waist circumference, and hip circumference, children exposed to 3+ HRFs had significantly higher AGE levels than those who did not have any HRF (OR: 9.21, 95% CI = 4.74, 17.91). These datasets are shown in Table 5 .

The statistical results of the EFA of HRFs in the training dataset are shown in Table 6 . All factors had good discriminant validity since their correlations were lower than 0.85.

In Table 7 , all HRF patterns were statistically significant with the BMI of participants by gender. Among males, BMI was taken as the dependent variable, with medium risk (OR: 2.81, 95% CI = 1.82, 4.34) and high risk (OR: 2.06, 95% CI = 1.05, 4.06) of overweight. Among females, BMI was taken as the dependent variable, with medium risk (OR: 0.90, 95% CI = 0.39, 2.08) and high risk (OR: 6.02, 95% CI = 2.42,14.99) of obesity. The results of adjusting co-variables are still significant, as shown in Table 7 .

Moderation analyses were performed with waist circumference and hip circumference as the control variables. The results are presented in Table 8 . First, HRFs significantly predicted the severity of BMI (β = -0.13), and gender was not associated with BMI (β = -0.065). Secondly, HRFs significantly predicted the severity of BMI (β = 0.29), and gender was also associated with BMI (β = 0.56). There were also a significant effect between gender and HRFs on AGE (β = -0.26) and BMI (β = 0.073).

As shown in Table 9 , there was also a significant three-way interaction among HRFs, BMI, and gender for AGE accumulation in the total sample (β = -0.30); after controlling for waist circumference and hip circumference, these results were also significant (β = -0.31).

In this study, non-invasive skin AGE levels were detected in community populations to observe possible influencing factors. The first is the detection of biochemical indicators, including SBP, TC, TG, HDL, BMI, HbA1c, FBG, and AGE. Normally, the generation of AGEs goes through a slow glycosylation stage, and hyperglycemia, hyperlipidemia, or oxidative stress can accelerate the accumulation of AGEs in the body (32). Obesity and what is referred to as insulin resistance (IR) syndromes, including hypertension, glucose intolerance, hyperinsulinemia, and hyperlipidemia, are common factors (33). In this study, multivariate logistic regression analysis showed that AGEs were associated with both T2DM and prediabetes risk. Therefore, it could be speculated that elevated levels of AGEs in the skin could be a warning sign of T2DM, and high serum glycated hemoglobin A1c (HbA1c) can accelerate the formation of AGEs and hydroxymethyl lysine. Even if HbA1c is well controlled, high levels of AGEs can also increase the risk of diabetic microvascular complications and form the “metabolic memory” effect of hyperglycemia (34). Patients with type 1 T2DM were followed up for 7 years. It was found that AGE fluorescence spectrum detection of skin could predict the occurrence of major adverse cardiovascular events in patients with type 1 T2DM; the higher the intensity of AGE fluorescence in skin, the higher the probability of myocardial infarction, stroke, lipid metabolism index, lower-limb paraplegia, or revascularization surgery in patients with type 1 T2DM (35). In addition, a previous study proposed that AGE was also connected to the pathophysiology of obesity (36). Another study has reported that obese adults (average body mass index [BMI] 33.2 kg/m²) exhibited significantly higher-circulating AGEs compared to overweight participants (average BMI 26.3 kg/m²) (37). Serum AGEs were found to correlate with triglyceride levels, and those with the highest AGEs had the most adverse lipid profiles (38). Another study demonstrated that AGE was also associated with kidney function (39). Similarly, age and gender differences were also verified to be associated with AGE. Serum AGEs have been reported to be in the range of 8.5 ± 0.9 units/ml in men less than age 45, 9.9 ± 1.5 in men older than 60, 7.9 ± 0.7 in women less than age 45, and 10.7 ± 1.1 in women older than 60 in a study of 172 healthy individuals (40).

Our research suggests that participants, especially those with a higher BMI and higher health risk factors, seem to be a vulnerable group for AGE accumulation. From a developmental disease perspective, obesity interacts with health risk factors, influencing the development of AGE accumulation. BMI, as a moderating variable, has gender-specific effects on the association between HRFs and AGE accumulation. The quality of gender moderates the relationship between BMI and AGE accumulation. As we all know, HbA1c and obesity were all associated with AGE accumulation; AGEs may play a more prominent role in the health of obese individuals, so obesity may be one of the health risk factors and pathophysiology mechanisms leading to AGE accumulation in health risk factors (39, 41). Moreover, we also added gender: according to the diathesis-stress model, females are more vulnerable to the higher effects of HRFs in a higher BMI (obesity), consequently leading to greater female-to-AGE accumulation. Although components demonstrating this inverse relationship included body mass index, TG, HbA1c, and insulin resistance, which were all inversely linked with RAGEs (receptor for AGEs) in both diabetic and non-diabetic participants (42), this also verifies that the indicators selected in this study are reasonable. Another reason was that AGEs increase with increasing BMI whereas the link is opposite for sRAGE; sRAGE may reflect tissue RAGE expression and that sRAGE may increase along with AGE in order to mount a counter-defense (43).

Fluorescent intensity of AGEs in the skin can reflect the longer time of blood glucose control, which may be one of the indicators to predict the chronic complications and mortality of T2DM (44). Previous studies have shown that the interaction between AGEs and its receptor RAGE causes oxidative stress, inflammation, and fibrosis, which can lead to endothelial dysfunction, atherosclerosis, vascular stiffness, progressive changes in renal structure, and impaired renal function (32, 45, 46). Therefore, non-invasive skin detection for AGEs is of great significance for the prevention and monitoring of T2DM and its complications. Venous blood FPG, OGTT, and glycated hemoglobin are important indicators for the screening and diagnosis of T2DM, but there are some problems, such as invasiveness, high cost, and long time to transport samples (47, 48). A large number of foreign studies have reported that skin autofluorescence can be used for T2DM screening, and the sensitivity of the diagnosis of T2DM is higher than FPG or HbA1c (48–50). Moreover, these non-invasive measurements of AGE accumulation may be considered as promising biomarkers of late diabetic complications.

The following are the study’s limitation: a large number of foreign studies have reported that skin autofluorescence (SAF) can be used for T2DM screening, and the sensitivity of the diagnosis of T2DM is higher than FPG or HbA1c. The authors did not investigate the subjects’ family history of T2DM, dietary habits, etc., which may affect the results of the study (39). In addition, Noordzij reported that skin autofluorescence detection may have some limitations, such as skin pigmentation, use of cream and sunscreen, and extreme congestion or vasoconstriction, which may affect the measurement results (51).