Terminal trajectory of HbA1c for 10 years supports the HbA1c paradox: a longitudinal study using Health and Retirement Study data

Objectives We aimed to assess the potential time-varying associations between HbA1c and mortality, as well as the terminal trajectory of HbA1c in the elderly to reveal the underlying mechanisms. Design The design is a longitudinal study using data from the Health and Retirement Study. Setting and participants Data were from the Health and Retirement Study. A total of 10,408 participants aged ≥50 years with available HbA1c measurements at baseline (2006/2008) were included. Methods Longitudinal HbA1c measured at 2010/2012 and 2014/2016 were collected. HbA1c values measured three times for their associations with all-cause mortality were assessed using Cox regression and restricted cubic splines. HbA1c terminal trajectories over 10 years before death were analyzed using linear mixed-effect models with a backward time scale. Results Women constitute 59.6% of the participants with a mean age of 69 years, with 3,070 decedents during the follow-up (8.9 years). The mortality rate during follow-up was 29.5%. Increased mortality risk became insignificant for the highest quartile of HbA1c compared to the third quartile (aHR 1.148, 1.302, and 1.069 for a follow-up of 8.9, 6.5, and 3.2 years, respectively) with a shorter follow-up, while it became higher for the lowest quartile of HbA1c (aHR 0.986, 1.068, and 1.439 for a follow-up of 8.9, 6.5, and 3.2 years, respectively). Accordingly, for both decedents with and without diabetes, an initial increase in HbA1c was followed by an accelerating terminal decline starting 5–6 years before death. Conclusions and implications The time-varying association between HbA1c and mortality mapped to the terminal trajectory in HbA1c. High and low HbA1c may have different clinical relationships with mortality. The HbA1c paradox may be partially explained by reverse causation, namely, early manifestation of death.


Introduction
The average plasma glucose level for the 2 to 3 months prior was reflected by glycated hemoglobin (HbA 1c ) (1), which increases with age (2)(3)(4).It is acknowledged that a high HbA 1c level is an independent indicator of mortality among middle-aged and older adults with and without diabetes (5,6).Nevertheless, studies demonstrating the detrimental effects of low HbA 1c and the protective benefit of high HbA 1c have emerged over the past few decades, which is counterintuitive and known as the HbA 1c paradox.This inconsistency regarding the association between HbA 1c and mortality was noteworthy among older adults (7)(8)(9).According to a recent largescale observational study conducted in primary care in the UK, this inconsistency may be interpreted in terms of possible varying magnitudes of the association between HbA 1c and mortality over time (8).Indeed, the effect of HbA 1c on mortality may vary across follow-up (10), which is seldom considered in prior studies.
Several researchers have highlighted the potential of reverse causation for explaining the varied effects of HbA 1c on mortality, or sometimes known as the HbA 1c paradox.This suggests that rather than being randomly associated with long-term mortality, low HbA 1c may instead be associated with poor health in older adults (11,12), which could be an early sign of mortality.However, such hypothesis has not been examined using formal analyses.Terminal trajectories with a backward time scale (13) can be used to characterize the exact changes of HbA 1c in the years immediately before death, and therefore can be applied to examine the early manifestations of death.Furthermore, several studies that have modeled biomarker terminal trajectories as time to death in older adults have identified a terminal decline for blood pressure (BP) (14-16) and total cholesterol (TC) (17), indicating that low BP or TC is a proxy of mortality.Nevertheless, the terminal trajectory of HbA 1c with proximity to death has not been assessed.Addressing such an issue will not only shed light on the understanding of the time-varying association of HbA 1c with mortality from a clinical perspective, but also inform risk assessment and treatment decision-making.
Therefore, the aims of this study were (1) to investigate the associations between HbA 1c measured at multiple times and mortality to assess the potential time-varying association, and (2) to estimate the terminal HbA 1c trajectories as death approached and compare them with the HbA 1c trajectory for survivors over a 10-year period in a middle-aged and older population with and without diabetes.

Study population
This study used data from the Health and Retirement Study (HRS).HRS is a nationally representative longitudinal survey of noninstitutionalized older adults aged ≥50 years in the United States (18).Biomarker measures including HbA 1c are examined during the enhanced face-to-face interview, with a randomized half of the sample first measured in 2006 and the other half measured in 2008.Each group then had follow-up exams every 4 years following the initial exam.For this study, we used data covering 2006 over 2016.The study population (n = 10,852) consisted of respondents who satisfied the following criteria: (1) completed the interview in 2006 or 2008 (i.e., baseline); (2) had baseline HbA 1c measures; and (3) had reported their history of diabetes (yes/no) at baseline.We further excluded participants who had missing data for any of the baseline covariates (n = 444).A total of 23,273 interviews of 10,408 participants were eligible (mean, 2.2 interviews/person) (Figure S1).Of those with baseline HbA 1c measure, 7,286 (70.0%) participated during the second wave (2010/ 2012) and 5,801 (55.7%) participated during the third wave (2014/ 2016).Having baseline diabetes was defined as individuals who answered "yes" to the question "Has a doctor ever told you that you have diabetes or high blood sugar?" or had baseline HbA 1c ≥6.5% (19).
In HRS, proxy interview rates across years (i.e., 2006-2016) ranged from 4.5% to 6.8%.All participants or their proxy respondents have provided written informed consent.
HbA 1c measurement (2006/2008, 2010/ 2012, 2014/2016) We used three waves of HbA 1c data collected at interviews of 2006/2008, 2010/2012, and 2014/2016 for analysis.In the HRS, HbA 1c was measured by dried blood spots (DBSs), which are highly correlated with the whole blood (r = 0.956) (20).Because the biomarker values based on DBS vary across assays and laboratories, the HRS adjusted DBS values to levels consistent with the National Health and Nutrition Examination Survey (NHANES) (21).We used the NHANES-equivalent assays for analysis, which are recommended by HRS and used in previous studies.

Mortality
Deaths from any cause through 2018 and the date of death (year) were determined by exit interviews with proxy respondents as well as the National Death Index data linked to respondents.For this study, included participants who died from 2006 to 2018 are identified as decedents, and those who survived 2018 were identified as survivors.

Covariates
Information on covariates were obtained from the baseline interview, including sociodemographic variables (sex, age, race, education, and marital status), lifestyle variables [physical activity, smoking, drinking, and body mass index (BMI)], and health conditions (diabetes, hypertension, heart diseases, and multimorbidity score).Details were presented in the Supplementary Materials (eMethod 1).

Statistical analyses
Baseline characteristics of participants by survival status at the end of follow-up were described.Comparisons between decedents and survivors were performed using t-test, Kruskal-Wallis test, or chi-square test.
Three Cox proportional regression models were used to examine the time-varying association between HbA 1c and mortality risk.HbA 1c values measured in 2006/2008, 2010/2012, and 2014/2016 were, respectively, the independent variables in the three models.Follow-up was from HbA 1c measurements until death or December 2018.As a result, the time-varying connection between HbA 1c and mortality can be examined under various follow-up durations (as shown in Figure S2).In particular, the quartiles of the baseline HbA 1c test were used to classify the HbA 1c values obtained in the three waves.We reported the hazard ratios (HRs) and 95% confidence intervals (CIs) for different HbA 1c categories in the three models respectively.Based on earlier research suggesting that 5.6%-6.5% may be the ideal ranges for overall survival, we selected the third quartile as our reference (22).Proportional hazards assumption was tested by the Schoenfeld residuals trend test (all p > 0.333).Analyses were adjusted for sociodemographic variables (model 1), additionally for lifestyle variables (model 2), and then for health condition variables (model 3).Restricted cubic splines (RCS) with four knots (selected based on model R 2 ) were applied to visualize the nonlinear association between continuous HbA 1c and mortality across different exposure times.
We estimated the terminal trajectory of HbA 1c using linear mixed models.Models were fitted with HbA 1c as the dependent variable, and survival status, time terms, and their interactions as independent variables.As for the time terms, we adopted a backward time scale spanning to 10 years before death or end of follow-up.The random effects for the intercept and time in the linear mixed models allowed for differences in HbA 1c at the intercept (time 0) and change in HbA 1c over time.Models were adjusted for age at death and covariates mentioned above.For better comprehension, continuous variables were centered around the baseline mean.Because of the non-normal distribution, the HbA 1c values were converted logarithmically, and the coefficients can be understood as percentage differences in means.
We performed subgroup analysis by diabetes status at baseline in all of the above analyses.For sensitivity analyses, we additionally adjusted for self-reported medication use relating to hypertension, diabetes, and heart diseases at baseline.Moreover, to control for hemoglobinopathy, we excluded participants with HbA 1c < 4.5% (n = 99).We also excluded those who had a diagnosis of cancer at baseline (n = 1701).Finally, we included participants who had missing covariates (n = 444) and re-analyzed by imputing the missing values.Statistical significance was set to two-sided p < 0.05.STATA version 14 and R version 4.1.2were used for analyses (23,24).Details about statistical analyses are provided in the Supplementary Materials (eMethod 2).

Results
Of the 10,408 included participants, 3,070 were identified as decedents while the remaining 7,338 were identified as survivors with a 5-year follow-up mortality rate of 13.3% and 29.5%, respectively, during the full follow-up.Table 1 displays the baseline characteristics by survival status.Decedents were older at baseline (76.3 vs. 65.9years) and time 0 (81.2 vs. 74.1 years), and had a higher proportion of men (47.0% vs. 37.6%) and self-reported history of diabetes at baseline [27.2% (n = 836) vs. 17.7% (n = 1,301)].The median level of HbA 1c at baseline was 5.8% in decedents and 5.6% in survivors.There were also differences between decedents and survivors in other sociodemographic, lifestyle, and health condition variables.Respondents who were excluded from the study were older, had more comorbidities, and had higher HbA 1c (Table S1).

Time-to-event analyses
We observed time-varying associations of HbA 1c with mortality across different follow-up durations (Table 2).For the Cox model with 2006/2008 HbA 1c measures as the independent variable (mean follow-up, 8.9 years), the highest quartile showed a significantly higher risk of mortality compared with the third quartile in all adjusted models (fully adjusted HR, 1.148; p = 0.013).This was also the case for the model of 2010/2012 (mean follow-up, 6.5 years; fully adjusted HR, 1.302; p = 0.001), but not for the model with a short follow-up (i.e., measures at 2014/2016; mean follow-up, 3.2 years; fully adjusted HR, 1.069; p = 0.597).Conversely, the lowest quartile of HbA 1c became significantly associated with a 43.9% higher mortality risk relative to the third quartile (p = 0.006) as followup was shorter (e.g., measures at 2014/2016).These temporal relationships between HbA 1c and mortality were visualized using RCS (Figure 1), showing a shift from "J-shape" to "L-shape" as the HbA 1c measurements get closer to death.
When stratified by diabetes status reported at baseline (Table S2; Figure S3), participants without diabetes showed that the highest quartile of HbA 1c was consistently associated with higher mortality risk compared with the third quartile across the three measurements.For participants with diabetes, there was a more significant increase in mortality risk for the lowest quartile of HbA 1c as follow-up was shorter.
Terminal HbA 1c trajectories over 10 years The clinical reasons underlying the time-varying link between HbA 1c and mortality were further suggested by the terminal HbA 1c trajectories for survivors and deceased individuals.As indicated (Figure 2A), survivors experienced a consistent increase in HbA 1c .In contrast, decedents initially experienced a higher level of and a steeper rise in HbA 1c , peaking in years 5-6, and then declined more quickly as they approached death.An intersection point was captured at year 5, whereby survivors had a higher HbA 1c level than decedents.Table 3 summarizes the intercept and slope differences of HbA 1c trajectories across survival status.There were significant interactions of survival status with linear time (p < 0.001) and quadratic time (p < 0.001), indicating an intensified decline in HbA 1c for decedents compared with survivors (Table 3).
Participants were further stratified by diabetes status collected at baseline.Those without diabetes followed a similar pattern of HbA 1c trajectories as observed in the total sample, although the overall HbA 1c levels were lower and changes were gentler (Figure 2B; Table 3).As for participants with diabetes (Figure 2C; Table 3), decedents showed a more remarkable upward trend than survivors until years 5-6, with a later intersection (at year 4.5) and a subsequently wider negative gap (i.e., decedents had lower HbA 1c levels) (Table S3).

Sensitivity analyses
Sensitivity analyses as described above did not alter the results (Tables S4-S10).

Discussion
This study shows that low HbA 1c is associated with higher mortality within short follow-up periods, while high HbA 1c is associated with higher mortality in long follow-up periods, indicating a change from a "J-shaped" to an "L-shaped" association between HbA 1c level and mortality as death † HbA 1c was categorized according to the quartiles of HbA 1c measurements at baseline, i.e., 1st quartile (HbA 1c ≤ 5.34%), 2nd quartile (5.34% < HbA 1c ≤5.69%), 3rd quartile (5.69% < HbA 1c ≤ 6.14%), and 4th quartile (HbA 1c > 6.14%).We treated the third quartile (5.69% < HbA 1c ≤ 6.14%) as the reference in all the three Cox models.‡ For HbA 1c measures at 2006/2008, there was a mean follow-up of 8.9 years (SD 2.5), with 92,188 person-years and deaths/total of 3,070/10,408.§ For HbA 1c measures at 2010/2012, there was a mean follow-up of 6.5 years (SD 1.8), with 47,622 person-years and deaths/total of 1,455/7,286.¶ For HbA 1c measures at 2014/2016, there was a mean follow-up of 3.2 years (SD 1.0), with 18,345 person-years and deaths/total of 520/5,801.approached.Accordingly, HbA 1c trajectory first rose and then terminally declined at an accelerated rate as death approached.

A B C
Participants who had and those who did not have diabetes at baseline followed a similar pattern.These findings revealed that the effect of HbA 1c on mortality that varied across follow-up in older adults may be a reflection of terminal HbA 1c trajectory.As a result, there may be differences in the clinical and biological significance of high and low HbA 1c levels with respect to mortality risk.Reverse causality provides some explanation for the HbA 1c paradox.

Comparisons with previous studies
Recently, the paradoxically protective effects of higher glycemia have been extensively investigated, with studies showing that HbA 1c <5.0% and ≥6.5% would increase 30-day, 90-day, and 1-year mortality for critically ill patients, as well as an approximate Ushape association between HbA 1c and the risk of mortality being recognized (25)(26)(27)(28).As for long-term effects, similar resultssubjects with HbA 1c level <6% and ≥10% were more likely to have in-hospitality mortality-were found (29).However, prior studies on the HbA 1c paradox have largely focused on mean HRs over the follow-up period, potentially ignoring the varied effects for the temporal pattern of exposure.A meta-analysis found that both high and low HbA 1c were risk factors for mortality, with duration of follow-up being one of the sources of heterogeneity (30).Accounting for this, Laiteerapong et al. reported that longer period of exposure to high HbA 1c (>8.0%) among adults with diabetes was linked to higher mortality risk (31).Comparably, a recent nationally representative study conducted in the United States showed that a low HbA 1c was associated with an elevated risk of all-cause death at 5 and 10 years of follow-up, respectively, of 30% and 12% (32).Although they involve different populations and HbA 1c levels, these findings were in parallel to our results, in which increased mortality risk became lower for high HbA 1c , while it became higher for low HbA 1c when follow-up was shorter (i.e., closer to death).Such results highlight the necessity of taking into account differences in magnitude when evaluating the effect of HbA 1c on mortality.In particular, when stratified by diabetes status, we observed that increased mortality risk was consistently correlated to high HbA 1c for participants without diabetes, while it was more significantly correlated to low HbA 1c for participants with diabetes.Therefore, while strict glycemic control may not always assist those with diabetes, maintaining a healthy HbA 1c level over time may be crucial for those without the disease (33).
A growing number of studies argue that the observed increased mortality risk associated with low HbA 1c or intensive glucose therapy may not be a casual effect of low HbA 1c (32,34); an epidemiological investigation using the United States national database also supported the idea that low HbA 1c level was a proxy of end stage of life after adjusting for an extensive set of potential confounders with flexible modeling (8).However, the above studies did not conduct a formal analysis of terminal HbA 1c trajectory in decedents.Our study found an accelerated terminal decline in HbA 1c over 5-6 years prior to death.Comparing survivors and decedents revealed that this terminal decline in HbA 1c might not be the result of aging.Furthermore, HbA 1c -lowering treatment also did not seem to explain such a decline since participants without diabetes also experienced a terminal decline.Therefore, our results expanded earlier studies on terminal trajectories of cardiometabolic indicators (14, 15, 17) and provide straightforward evidence for the hypothesis that low HbA 1c is a sign of mortality.Although the exact mechanisms causing the observed terminal HbA 1c fall are unknown, they may have to do with failing organs, malnourishment, and unfavorable profiles of components associated with red blood cells, among others (11,35).Indeed, the relationship between low HbA 1c and mortality was lessened in our study's sensitivity analyses by removing people with an HbA 1c of less than 4.5% (probable hemoglobinopathy).
It is interesting to note that the decedents' trajectory showed an early rise in HbA 1c .Some post-trial analyses have presented the HbA 1c "legacy effect"; that is, exposure to adverse glucose control had a prolonged influence on the risk of future mortality (36).
Several longitudinal studies have shown the adverse effects of poor glycemic control on major cardiovascular events and immune function related to COVID-19 over the years (37,38).These data indicated that high HbA 1c level may be a predictor for the long-term mortality risk.Furthermore, the initially rising and subsequently falling HbA 1c trajectory is exactly mapped to the time-varying associations between HbA 1c and mortality across follow-up, offering a mechanism explanation for the time-varying association.In fact, our conclusion is in line with several clinical mechanism studies that show that early glycemic intervention targeting glucose sodium-cotransporter-2 inhibitors can improve myocardial function (39)(40)(41).Conjunctively, we supposed that high HbA 1c may be a long-term predictor of death, whereas low HbA 1c is more likely a reverse causation, namely, an early manifestation of mortality, for this middle-aged and older adult population.

Meaning of findings
This study will be beneficial for the understanding of physiological change before death and the identification of risk populations.Results highlighted the importance of long-term HbA 1c monitoring as early as possible in light of the probability of notable alterations occurring 10 years before to mortality (42,43).In a clinical setting, effective glycemic treatment can be introduced early to patients with increased HbA 1c (44,45).However, an intensified glycemic control may not be appropriate for those who are at the end of life, supporting recent recommendations that the glycemic management of older adults with diabetes should be individualized depending on the patient's life expectancy and overall health status (46,47).Moreover, the identification of turning point at 5 years before death could have implications for the less glycemic targets in the context of life span.This study also supported the rationale of "reverse epidemiology/reverse causation" (48); that is, patients with a lower HbA 1c were generally closer to death, which may partly explain the paradoxical effects of HbA 1c in previous studies.

Strengths and limitations
This study allowed for additional insight into the processes behind the time-varying effects of HbA 1c on mortality by combining retrospective terminal trajectory studies with prospective time-toevent analysis.Moreover, to our knowledge, this study provides the first evidence of the terminal trajectory in HbA 1c using a backward time scale for 10 years before death.Another advantage is the application of a nationally representative sample of middle-aged and older adults drawn from the HRS panel study.However, some limitations should be considered.Firstly, we are unable to examine the HbA 1c trajectory among specific causes of death because of a lack of information on cause-specific mortality.In addition, it remains unclear to what extent the diabetes phenotypes, treatments, and disease duration impact the HbA 1c terminal trajectory, although we managed to control the usage of medication in sensitivity analyses.There was evidence that patients with a long duration of diabetes were more inclined to have poor outcomes when exposed to low HbA 1c than those with a short duration (49).Therefore, future studies are expected to validate the present findings accounting for these factors.Moreover, we cannot completely rule out misclassification, nor did we take time-varying confounders into account.Information reported by agents at the time of data collection was incorporated into HRS, and this study may be exposed to additional unmeasured confounders.Moreover, we adopted the logarithmic transformation of HbA 1c data as the analytical variable in our study, which may have limited how our findings could be interpreted and directly compared to those of other investigations.Furthermore, contrary to what previous studies have found, our study did not identify a steeper decline in HbA 1c during the last 2 years before death (14,17).This may arise from the relatively long measurement interval.Nevertheless, obtaining frequent HbA 1c in representative populations over 10 years before death can be challenging, and the present data with repeated measurements are qualified for analysis.

Conclusions and implications
This analysis showed time-varying associations between HbA 1c and mortality across different durations of follow-up, shifting from "J-shaped" to "L-shaped" as death approached.Such associations are exactly mapped to the HbA 1c terminal trajectory showing an early increase followed by an accelerating terminal decline for 5-6 years before death in HbA 1c , which was not observed in survivors.These results suggested that there might be differences in the clinical significance of high and low HbA 1c with regard to mortality: high HbA 1c may be a long-term predictor of death, while low HbA 1c may be an early manifestation of death.Thus, the HbA 1c conundrum in earlier research may be partially explained by reverse causation.Our findings underscore the necessity of long-term HbA 1c monitoring as early as possible.Additionally, effective glycemic control should be introduced early but may not be appropriate when patients are at the end of life.

FIGURE 1 2
FIGURE 1 Nonlinear association of HbA 1c measured in 2006/2008 (A), 2010/2012 (B), and 2014/2016 (C) with mortality for total sample.Data were derived from the Cox models with restricted cubic splines.Analyses were adjusted for HbA 1c ; sex; age; race; marital status; education; physical activity; smoking; drinking; body mass index; history of hypertension, diabetes, and heart diseases; and multimorbidity score.HbA 1c , glycated hemoglobin; HR, hazard ratio; CI, confidence interval.

TABLE 1
Baseline characteristics (2006/2008) of participants by survival status at the end of the follow-up (2018).
* Heart diseases included heart attack, coronary heart disease, angina, and congestive heart failure.†Multimorbidity score was created as the count of nine chronic diseases: hypertension, diabetes, heart diseases, lung diseases, stroke, cancer, psychiatric problems, dementia/Alzheimer's disease, and arthritis.‡Time 0 was 2018 for survivors and date of death for participants who died between baseline (2006/2008) and 2018.

TABLE 2
Association between multiple measurements of HbA 1c and subsequent mortality.
HbA 1c , glycated hemoglobin; CI, confidence interval.* Model 1 adjusted for sex, age, race, marital status, and education; Model 2 additionally adjusted for physical activity, smoking, drinking, and BMI; Model 3 additionally adjusted for history of hypertension, diabetes and heart diseases, and multimorbidity score.

TABLE 3
Coefficient estimates of HbA 1c trajectories for decedents and survivors for 10 years from linear mixed-effect models.Data presented in the table were derived from the linear mixed-effect models using the log of HbA 1c values.Models were adjusted for sex, age, race, marital status, education, physical activity, smoking, drinking, body mass index, history of hypertension, diabetes and heart diseases, and multimorbidity score.CI, confidence interval.† Survival status at 2018, including decedents (n = 3,070) and survivors (n = 7,338).‡ Time was the years before death or end of follow-up.Time 0 was 2018 for survivors and date of death for participants who died between baseline (2006/2008) and 2018.