Participants with T2DM from 14 communities in Beijing were recruited from Beijing Community Diabetes Study between August 2008 and July 2009, which has been described in detail previously (12). The participants enrolled in the study were followed up for 10 years (2008–2018). On account of most of the clinical outcomes (76%) that occurred after 2014, we divided the 10-year follow-up into two stages for the present data analysis to explore the effect of HbA1c trajectories on future clinical outcomes. Stage I (from 2008 to 2014) was used to identify the HbA1c trajectories and to calculate the HbA1c-SD and HbA1c-CV. Stage II (from 2014 to 2018) was used to collect the records of the occurrence of future diabetes-related clinical outcomes.

T2DM was defined according to the criteria of the WHO (1999) (13). A total of 4,525 T2DM participants without hemoglobin disease, anemia, or severe renal dysfunction were included in the initial study, while those with missing data of outcomes or who did not finish the 10-year follow-up were excluded (n = 630). Participants who had less than three measurements of HbA1c or with the clinical outcomes that occurred at stage I (from 2008 to 2014) were also excluded (n = 1,734). A final total of 2,161 participants were included ( Figure 1 ).

All the T2DM patients were managed by the community doctors and specialists from Beijing Tongren Hospital. Patients were given health education, including individualized guidance on reasonable diet and exercise. During the annual follow-up, questionnaire survey, physical examination, and the biochemical indexes evaluation were conducted, and electrocardiogram and eye examination were also performed simultaneously.

The date of birth, gender, duration of diabetes, concomitant diseases, smoking habits, income, education, and medical history at baseline were collected using a questionnaire by trained interviewers. The income was divided into three categories: less than 2,000, 2,000 to 4,000, and over 4,000 yuan per month. Participant education was categorized into three groups, namely, primary education (illiterate and primary school graduates), secondary education (junior and senior high school graduates), and higher education (college graduates and above). Anti-hyperglycemic therapy was classified as non-drug therapy, oral anti-hyperglycemic drugs, insulin treatment, and oral anti-hyperglycemic drugs combined with insulin. Blood pressure, body mass index (BMI), and waist circumference (WC) were also collected using standard methods. Fasting blood samples were collected for measurements of HbA1c, blood glucose, serum creatinine, serum uric acid, alanine aminotransferase, and lipids. HbA1c was measured using a Bio-Rad Variant hemoglobin analyzer by a central endocrinology laboratory in Beijing Tongren Hospital.

Outcomes of interest included all-cause death, cardiovascular events (acute myocardial infarction, coronary artery bypass grafting, coronary stent, unstable angina pectoris, hospitalization for heart failure, and installation of cardiac pacemaker), cerebrovascular events (cerebral hemorrhage, cerebral infarction, transient ischemic attack, and subarachnoid hemorrhage), renal events (new-onset proteinuria, microalbuminuria turning into macroalbuminuria, doubling of the serum creatinine level, and dialysis), diabetic eye disease events (fundus laser and vitrectomy treatment), and composite endpoint events (all-cause death, cardiovascular events, cerebrovascular events, renal events, and diabetic eye disease events), which all occurred from 2014 to 2018. All outcomes were adjudicated by an independent committee, which was in charge of assignments on validation of data and events based on the outcome criteria.

The HbA1c-SD and HbA1c-CV were calculated using HbA1c values measured in Stage I to evaluate HbA1c variability. As the number of HbA1c measurements would affect the SD value, the HbA1c-SD was adjusted for the number of HbA1c assessments according to the formula: adjusted SD of HbA1c ( HbA 1 c − adjSD ) = HbA 1 c − SD/ n / ( n − 1 ) (14). Furthermore, the HbA1c-CV was calculated as the HbA1c-SD to HbA1c–MEAN ratio to correct for larger SDs resulting from higher absolute HbA1c–MEAN values (14).

HbA1c trajectories were identified using the latent growth mixture model (LGMM) with Mplus 8.0 (11). We assumed that there were multiple potential growth trajectories in the population, and each potential trajectory represented a subclass that had different growth patterns. The model contains two kinds of latent variables, including continuous latent variables and classified latent variables. The former contains intercept and slope of growth characteristic parameters, while the latter refers to the mutually exclusive subgroups. We used potential subgroups in increasing order to fit linear and quadratic growth mixed models. The optimal model was selected through the information indexes, including the Lo–Mendell–Rubin fit index, bootstrapped likelihood ratio test, entropy, and the adjusted Bayesian information criterion. Each trajectory subgroup must be large enough to have clinical relevance, namely, containing a minimum of 1% of patients (11), and then we assigned patients to the subgroup where they had the highest posterior probability.

Normally distributed quantitative variables were expressed as the mean ± SD, whereas categorical variables were shown as numbers (percentages). Parameters that were not normally distributed were presented as the median [interquartile range (IQR)]. Comparison of variables among groups was performed using the Kruskal–Wallis and one-way ANOVA. Cox proportional hazards models were used to evaluate the relationship between HbA1c trajectories, HbA1c-CV, or HbA1c-adjSD, and future outcomes. SPSS (version 24.0) was used to perform all statistical analyses. p < 0.05 (two-sided) indicated statistical significance.

As shown in Figure 2 , three patterns of HbA1c trajectories were identified by LGMM, including the low stable (88.34%), pre-stable and post-increase (5.83%), and gradual decreasing (5.83%) HbA1c trajectories. The low stable group had a relatively good glycemic control with the lowest levels of HbA1c. The pre-stable and post-increase group had higher than recommended but stable HbA1c initially and significantly increased at the later stage, while those in the gradual decreasing group had the highest levels of HbA1c in the first few years but gradually decreased over time.

Baseline characteristics of three patterns of HbA1c trajectories groups are shown in Table 1 . The median baseline fasting plasma glucose (FPG) of individuals was 6.8 (IQR, 6.0–8.0), 8.4 (IQR, 7.0-10.9), and 11.2 mmol/L (IQR, 9.0-13.6), and the median baseline HbA1c was 6.7% (IQR, 6.2–7.4), 8.0% (IQR, 7.2-9.7), and 9.9% (IQR, 9.1-11.5) in the low stable, pre-stable and post-increase, and gradual decreasing groups, respectively. Although the glycemic control was significantly different among the three groups, there were no significant differences in gender, T2DM duration, anti-hyperglycemic therapy, high-density lipoprotein cholesterol (HDL-C), alanine aminotransferase, systolic blood pressure (SBP), smoking habits, and the prevalence of hypertension, stroke, and dyslipidemia at baseline. Compared with participants in the low stable group, those in the pre-stable and post-increase and gradual decreasing groups were more likely to be younger and had higher BMI, diastolic blood pressure (DBP), WC, triglycerides (TG), total cholesterol (TC), and low-density lipoprotein cholesterol (LDL-C) but had lower income and proportion of higher education. We also found that participants in the pre-stable and post-increase group had lower levels of uric acid, a lower proportion of angiotensin-converting enzyme inhibitors (ACEIs) or angiotensin II receptor blockers (ARBs) takers, and a lower prevalence of coronary heart disease at baseline as compared to those in the low stable group. In addition, participants in the pre-stable and post-increase group also had the lowest levels of serum creatinine among the three groups.

From 2014 to 2018, 96 cardiovascular events, 70 cerebrovascular events, 125 renal events, 47 diabetic eye disease events, 356 composite endpoint events, and 94 cases of all-cause death occurred among 2,161 individuals with T2DM. As shown in Figure 3 , compared to the low stable group, the incidence of renal events and composite endpoint events was significantly higher in the pre-stable and post-increase and gradual decreasing groups (both p < 0.001), and the prevalence of diabetic eye disease events was also significantly higher in the latter group (p < 0.05).

Cox proportional hazards models of the HbA1c trajectories with future clinical outcomes were shown in Table 2 . In unadjusted analyses (Model 1), both the pre-stable and post-increase and gradual decreasing groups had a higher risk of renal events than the low stable group (both p < 0.001). After adjustment for gender, age, duration of T2DM, BMI, smoking habits, baseline concomitant diseases, TG, LDL-C, blood pressure, anti-hyperglycemic therapy, and ACEI or ARB treatment (Model 2), the increased risk of renal events remained statistically significant for the pre-stable and post-increase [hazard ratio (HR) = 2.91, 95%CI: 1.60–5.31, p < 0.001] and gradual decreasing (HR = 2.44, 95%CI: 1.37–4.34, p = 0.002) groups, respectively. In further adjustment for average HbA1c, the main trend of the risk of renal events for the pre-stable and post-increase (HR = 2.83, 95%CI: 1.25–6.41, p = 0.013) and gradual decreasing (HR = 2.37, 95%CI: 1.08–5.21, p = 0.032) groups remained significant (Model 3). The gradual decreasing group also had a higher risk of diabetic eye disease events (HR = 2.62, 95%CI: 1.10–6.22, p = 0.029) and composite endpoint events (HR = 2.00, 95%CI: 1.41–2.83, p < 0.001) than the low stable group in unadjusted analyses (Model 1). However, after adjustment for gender, age, duration of T2DM, BMI, smoking habits, baseline concomitant diseases, TG, LDL-C, blood pressure, anti-hyperglycemic therapy, ACEI or ARB treatment, and average HbA1c (Model 3), these differences were no longer significant (both p > 0.05). For macrovascular events, the association between cardiovascular events or cerebrovascular events and HbA1c trajectories did not reach a significant difference in the models without and with adjustments (all p > 0.05, Models 1 to 3). In addition, the pre-stable and post-increase group had a higher risk of all-cause death (HR = 3.02, 95%CI: 1.46–6.24 p = 0.003) and composite endpoint events (HR = 2.12, 95%CI: 1.42–3.15, p < 0.001) than the low stable group after adjustment for gender, age, duration of T2DM, BMI, smoking habits, baseline concomitant diseases, TG, LDL-C, blood pressure, anti-hyperglycemic therapy, and ACEI or ARB treatment (Model 2), and these associations remained statistically significant after additional adjustment for average HbA1c (both p < 0.05, Model 3).

We also explored the association of HbA1c-adjSD and HbA1c-CV quartiles with future clinical outcomes ( Tables 3 , 4 ). In unadjusted analysis (Model 1), using quartile1 of the HbA1c-adjSD as a reference, the HRs for renal events were 1.28 (95%CI: 0.68–2.40, p = 0.443), 2.28 (95%CI: 1.30–4.00, p = 0.004), and 2.89 (95%CI: 1.68–4.96, p < 0.001) for quartile2, quartile3, and quartile4, respectively, and p for trend<0.001 ( Table 3 ). After adjustment for gender, age, duration of T2DM, BMI, smoking habits, baseline concomitant diseases, TG, LDL-C, blood pressure, anti-hyperglycemic therapy, and ACEI or ARB treatment, this trend remained significant (p for trend = 0.001, Model 2). In further adjustment for average HbA1c, the HRs were 1.32 (95%CI: 0.67–2.59, p = 0.417), 2.40 (95%CI: 1.29–4.46, p = 0.006), and 2.40 (95%CI: 1.23–4.66, p = 0.010) for quartile2, quartile3, and quartile4, respectively, and p for trend = 0.014 (Model 3). We also found that the risk of composite endpoint events was significantly higher in the third quartile than the first quartile of the HbA1c-adjSD after adjustment for average HbA1c and other variables (HR = 1.53, 95%CI: 1.09–2.15, p = 0.014, Model 3, Table 3 ). Similarly, both univariate and multivariate Cox HR models (all p for trend <0.05, Models 1–3) indicated that participants in the fourth and third quartiles of the HbA1c-CV were at higher risk of renal events compared to participants in the first quartile ( Table 4 ). The risk of composite endpoint events was significantly higher in the third and fourth quartiles than in the first quartile of the HbA1c-CV after adjustment for gender, age, duration of T2DM, BMI, smoking habits, baseline concomitant diseases, TG, LDL-C, blood pressure, anti-hyperglycemic therapy, and ACEI or ARB treatment (p for trend = 0.012, Model 2); however, in further adjustment for average HbA1c, these associations were no longer significant (p for trend = 0.164, Model 3, Table 4 ).