Prospective transitions in hemoglobin A1c following gestational diabetes using multistate Markov models

Abstract

We characterized the state-to-state transitions in postpartum hemoglobin A1c levels after gestational diabetes, including remaining in a state of normoglycemia or transitions between prediabetes or diabetes states of varying severity. We used data from the APPLE Cohort, a postpartum population-based cohort of individuals with gestational diabetes between 2009 and 2011, and linked A1c data with up to 9 years of follow-up (n = 34 171). We examined maternal sociodemographic and perinatal characteristics as predictors of transitions in A1c progression using Markov multistate models. In the first year postpartum following gestational diabetes, 45.1% of people had no diabetes, 43.1% had prediabetes, 4.6% had controlled diabetes, and 7.2% had uncontrolled diabetes. Roughly two-thirds of individuals remained in the same state in the next year. Black individuals were more likely to transition from prediabetes to uncontrolled diabetes (adjusted hazard ratio [aHR] = 2.32; 95% CI, 1.21-4.47) than White persons. Perinatal risk factors were associated with disease progression and a lower likelihood of improvement. For example, hypertensive disorders of pregnancy were associated with a stronger transition (aHR = 2.06; 95% CI, 1.39-3.05) from prediabetes to uncontrolled diabetes. We illustrate factors associated with adverse transitions in incremental A1c stages and describe patient profiles that may warrant enhanced postpartum monitoring.

Introduction

Pregnancy is a transient stressor that can unmask or exacerbate existing cardiometabolic dysfunction in some predisposed individuals.^1,2 Gestational diabetes mellitus (GDM), glucose intolerance during pregnancy, is one such condition that can signal later-life type 2 diabetes and cardiometabolic dysfunction.^3-5 In New York City (NYC), an estimated 4% of White, 6% of Black and Latina, and 12% of Asian birthing people are diagnosed with gestational diabetes, potentially fueling later-life disparities.⁶ Meta-analyses have demonstrated that gestational diabetes is associated with a 10-fold and 20-fold increased risk of type 2 diabetes at 10 and 20 years postpartum, respectively.^4,5 In NYC, we found that a GDM diagnosis was associated with a 12-fold increased risk of diabetes over 9 years of follow-up.⁶ Among those with GDM, we documented stark racial-ethnic inequities in cumulative diabetes incidence.⁷ Yet despite evidence of an association between GDM and later-life diabetes, it is unclear how A1c trajectories change after a gestational diabetes pregnancy, including the likelihood of remaining in a nondiabetes state or movement to prediabetes or diabetes stages of varying severity or control.

A growing body of literature examines differences in the onset of diabetes after GDM by characteristics such as race/ethnicity, nativity, and health behaviors such as postpartum weight loss and breastfeeding.^8-12 For example, Black and Hispanic race/ethnicity has been associated with a stronger transition from GDM to later-life type 2 diabetes. Most research to date, however, does not describe how these characteristics predict adverse or favorable transitions at more incremental stages of A1c progression. Race/ethnicity and socioeconomic status, for example, in part represent the accumulation of “biological weathering” due to chronic stress, disadvantage, and discrimination over the life course, which is implicated in type 2 diabetes disparities.^13-15 Perinatal characteristics may also signal underlying metabolic and vascular dysfunction, which manifests with the additional demands of pregnancy.¹⁶ Such conditions have the potential to function as acute, easily ascertained indicators of postpartum risk. For example, characteristics in prepregnancy (prepregnancy obesity, chronic hypertension),¹⁷ pregnancy (excess gestational weight gain),¹⁸ and hypertensive disorders of pregnancy¹⁹ and birth (shoulder dystocia,²⁰ macrosomia,²¹ and preterm birth²²) are associated with later-life type 2 diabetes and/or cardiovascular risk.

Understanding the influence of sociodemographic and perinatal characteristics could help identify patient profiles associated with adverse or favorable transitions in postpartum hemoglobin A1c levels to inform targeted preventative or therapeutic intervention. However, to date, most predictors of diabetes incidence and control have focused on the transition between two A1c states, such as diseased and nondiseased, using traditional statistical methods such as generalized linear models and survival analysis.^3-5,9,23 Multistate Markov models are a novel application to the study hemoglobin A1c levels following gestational diabetes, which allow estimation of how individuals change between incremental disease states over time. This includes estimation of the probability and instantaneous rate of transitioning between disease stages and the predictors of transition intensities. Applying multistate methods to the study of postpartum A1c levels can further elucidate the natural course of disease progression as well as inform targeted touch points for intervention.

To address these gaps, we used 2009-2017 data from the APPLE Cohort, a population-based cohort of linked birth records, hospital discharge, and A1C Registry data in NYC.⁶ Our objective was to develop a model to represent bidirectional changes in hemoglobin A1c levels following gestational diabetes, including normoglycemia, prediabetes, controlled diabetes, and uncontrolled diabetes. We also describe the sociodemographic and perinatal characteristics as predictors of state-to-state transitions in diabetes risk and severity.

Methods

Data sources

Data from 2009 to 2017 were drawn from the APPLE Cohort, a retrospective population-based cohort of birthing individuals in NYC (n = 798 371). Vital statistics birth records were matched with (1) the NYC A1C Registry and (2) Statewide Planning and Research Cooperative System database to obtain hemoglobin A1c values, test dates, and hospital diagnosis codes at discharge from delivery between 2009 and 2017, respectively. Institutional review board approval was obtained by the NYC Department of Health and Mental Hygiene, the New York State Department of Health, and the Icahn School of Medicine at Mount Sinai.

Sample population

An initial sample of n = 64 212 people had gestational diabetes (birth record indication or hospital record International Classification of Diseases, Ninth Revision [ICD-9]: 648.01-648.04, International Classification of Diseases, Tenth Revision [ICD-10]: O24.4 codes) at index pregnancy (ie, the first pregnancy during study observation)²⁴ between 2009 and 2017. We arrived at a final sample size of n = 34 171 people after excluding those with a record of diabetes prior to the second trimester (birth record or ICD-9: 250.x, 648.0, ICD-10: E08-E13x, O24.0, O24.1, O24.3, O24.8, Z79.4 codes) (n = 1590) and those with fewer than 2 A1c tests in the A1C Registry following 12 weeks postpartum (n = 28 451).^24,25 The characteristics of participants with and without at least 2 postpartum A1c visits are presented in Appendix S1, Table S1. For all participants, observation began at 12 weeks postpartum to allow resolution of gestational diabetes and ended at their last A1c test date (censored date) or observation end (December 31, 2017).

Measures

Outcomes

The following A1c thresholds were used as a proxy for the following diabetes states in the absence of medical diagnoses: (1) no diabetes (A1c < 5.7%), (2) prediabetes (5.7 ≤ A1c < 6.5%), (3) controlled diabetes (6.5 ≤ A1c < 7.0%), (4) uncontrolled diabetes (A1c ≥ 7.0%), and (5) censored (no subsequent A1c values following last record in the Registry).²⁶ We note that data contained within the NYC A1C Registry do not allow a distinction between type 1 and 2 diabetes.²⁷ We henceforth refer to the risk of “diabetes” in reported results. However, we expect > 95% cases to be type 2 diabetes, as this is the most common form of diabetes among adults.²⁷

Sociodemographic characteristics

We obtained maternal sociodemographic characteristics from the birth certificate at the index pregnancy. These included age group (10-24, 25-34, and 35+ years), race/ethnicity (non-Hispanic Black [henceforth “Black”], Hispanic, non-Hispanic White [henceforth “White”], or non-Hispanic Asian [henceforth “Asian”]) as self-reported on the birth record,^28-30 nativity (foreign born vs US born), educational attainment (less than college degree vs college graduate or higher), public vs private insurance,^31,32 and year of delivery. We consider race/ethnicity to be a social construct shaped by social relations, including historical and present-day discrimination, and cultural practices rather than biological categories.³³

Perinatal characteristics

We examined perinatal characteristics across prepregnancy, pregnancy, and birth. Prepregnancy characteristics at index pregnancy included preexisting chronic hypertension (ICD-9: 401.x-405.x, 642.0x-642.2x; ICD-10: O10.x, O11.x; or birth record indication)³⁴ and prepregnancy obesity (body mass index [BMI] kg/m² > 30) (birth record).^35,36 Pregnancy characteristics included hypertensive disorders of pregnancy (gestational hypertension; ICD-9: 642.3x; ICD-10: O13.x; or birth record indication) or eclampsia (ICD-9: 642.4x-642.6x, 642.7x; ICD-10: O11.x, O14.0-O14.2, O149; or birth record indication)^37,38 and excess gestational weight gain (exceeding the amount recommended for prepregnancy BMI category or not).¹⁸ Adverse birth characteristics included macrosomia (birth weight greater than 4500 g via birth record),³⁰ infant shoulder dystocia (ICD-9: 660.4, 600.41, 600.43; ICD-10: O66), preterm birth (gestational age < 37 weeks or not),³⁰ and cesarean section (birth record).³⁹ Finally, we examined exclusive breastfeeding at hospital discharge (birth record).³¹

Analysis

We estimated the transition probabilities between diabetes states from an index gestational diabetes pregnancy over time using unadjusted and adjusted multistate Markov models.^40,41 We examined 5 states: no diabetes, prediabetes, controlled diabetes, uncontrolled diabetes, and censored, the final absorbing state. Possible paths are indicated by arrows (not all paths are possible). For example, to correspond with the clinical classification of patients, once a patient entered a diabetes state (controlled or uncontrolled), a reversion back to a predisease state (ie, prediabetes or no diabetes) was not permitted. Those with a history of diabetes whose A1c values dropped below the 6.5% threshold were considered to have “controlled diabetes.” Low probability transitions (< 0.10) were restricted (eg, from “no diabetes” to “uncontrolled” or “controlled” diabetes) due to inadequate sample size. The probability of transitioning to a censored state is reported in the tables but not discussed in the text since the reasons for censoring are unknown. Observations for variables with missing values were omitted from analysis.

Multistate Markov models describe the process in which an individual moves between a series of states in continuous time.^40,41 As data consist of observations at arbitrary times (eg, doctor’s visits), the exact times of when states change are unobserved. To maximize the potential that each participant contributed information to the model for an observable transition, state-to-state transitions were examined at bounded time intervals (12-month increments). Models estimate the instantaneous rate of transition between stages of disease as well as the probability of transitioning from one state to another at a given time juncture. This includes, for example, the likelihood that someone in a prediabetes stage at a given interval (i) transitioned to a no-diabetes or controlled or uncontrolled diabetes state during the next wave (i + 1). If a participant had multiple A1c tests within a given year, the first (of 2) or median (if 3 or more) test was selected. In this way, we assume noninformative sampling.⁴⁰ In the Supplemental Files, we include a sensitivity analysis that selects the highest value A1c test in the event of multiple tests in a given year. Multistate Markov models assume that transitions across diabetes states are a memoryless condition (ie, are informed by membership to the state immediately prior) rather than an individual’s complete history of prior intervals.^41,42 We assessed model fit as well as the reasonableness of the memoryless transition assumption using empirical tests⁴³ (see Appendix S2, Figures S1-S2, Table S3).

Next, we estimated the mean sojourn time, or average length of time (in years) an individual spent in each disease state before transitioning to another state with accompanying 95% confidence intervals (CIs).

We then incorporated sociodemographic predictors into the Markov model and examined transition-specific hazard ratios and 95% CIs. Estimated hazard ratios (HRs) represent each covariate effect on the intensity of transition at a given time juncture.⁴⁰ Finally, perinatal characteristics were examined in unadjusted models and those adjusted for sociodemographic characteristics, which were hypothesized to be more distal risk factors on the underlying causal pathway. We compared unadjusted estimates with those adjusted for sociodemographic factors.

We performed data management using SAS Enterprise (v7.1.5) and statistical analysis using RStudio (v1.6.9). Multistate analyses were performed using the “msm: Multi-state Markov Model” package version 1.6.7⁴⁰ and the “markovchain” version 0.8.6⁴⁴ packages.

Sensitivity analyses

We describe study results from the vantage point of a given gestational diabetes pregnancy. However, given we do not have information on whether pregnancies prior to the start of study observation were complicated by gestational diabetes, we conducted a sensitivity analysis to estimate the transition probabilities among those with no previous live births.

Results

Among individuals with a gestational diabetes pregnancy between 2009 and 2017 and no history of diabetes, 56% had at least 2 follow-up A1c tests after the 12 weeks following pregnancy (Table S1). Those without postpartum follow-up were, on average, more likely to be White, had a younger age distribution, were less likely to have multiple previous births, and were less likely to be overweight or obese than those screened (Table S1). Of those tested postpartum, the average number of follow-up tests was 3.2, with a median of 2.0 (interquartile range [IQR], 2.0, 5.0). The average number of 12-month intervals an individual contributed to before entering the “censored” state was 3.4, with a median of 3.0 (IQR, 2.0, 8.0). Of those screened, 40% had no diabetes (A1c < 5.7%) in the first year postpartum, 46% had prediabetes (5.7% ≤ A1c < 6.5%), 6% had controlled (newly diagnosed diabetes) (6.5% ≤ A1c < 7.0%), and 9% had uncontrolled (newly diagnosed) diabetes (A1c ≥ 7.0%) (Table 1). Those with uncontrolled diabetes at first postpartum assessment had, on average, lower educational attainment, higher parity, and higher incidence of adverse perinatal health outcomes relative to other disease stages.

Table 1.

Sociodemographic and clinical characteristics by diabetes state during the first year postpartum among persons with GDM at baseline, 2009-2017 (n = 34 171).

	No diabetes   (A1c < 5.7%)   n = 15 425	Prediabetes   (5.7 ≤ A1c < 6.5%)   n = 14 729	Controlled diabetes (6.5 ≤ A1c < 7%)   n = 1565	Uncontrolled diabetes (A1c > 7.0%)   n = 2452
State prevalence	45.1	43.10	4.58	7.20
Age group
10-24 years	11.85	8.28	7.01	7.80
15-24 years	56.42	51.67	49.53	54.74
35+ years	31.73	40.05	43.46	37.46
Race/ethnicity
Black	16.09	22.42	30.12	29.75
Hispanic	29.11	30.32	31.81	39.4
White	22.11	11.02	6.94	6.87
Asian	31.26	34.4	28.37	22.12
Other/unknown	1.48	1.91	2.14	2.39
Nativity
US born	33.11	28.09	27.43	32.83
Foreign born	66.89	71.91	72.57	67.17
Insurance
Public or none	63.77	68.88	72.64	73.6
Private	35.49	30.44	26.75	25.65
Missing	0.74	0.69	0.62	0.77
Educational attainment
College or higher	55.31	50.51	44.28	39.89
Less than college degree	44.40	49.15	55.34	59.83
Missing	0.29	0.35	0.38	0.29
Subsequent births
1	64.54	70.17	68.94	70.84
2	29.86	25.09	21.29	19.81
3+	5.61	4.74	9.77	9.36
Previous live births
0	53.39	43.09	35.58	33.67
1	25.38	28.72	28.77	27.31
2+	8.45	11.96	16.11	17.33
Prepregnancy obesity
No	77.64	65.60	50.10	47.19
Yes	21.58	33.4	48.43	51.31
Missing	0.78	0.92	1.47	1.51
Excessive gestational weight gain
No	61.37	58.78	54.89	52.76
Yes	38.63	41.22	45.11	47.24
Preexisting hypertension
No	97.01	94.94	88.34	87.4
Yes	2.99	5.06	11.66	12.6
Preterm birth
No	88.95	86.44	83.6	79.52
Yes	11.05	13.56	16.4	20.48
Hypertensive disorders of pregnancy
No	90.25	87.81	79.45	75.94
Yes	9.75	12.19	20.55	24.06
Cesarean delivery
No	58.27	51.98	45.22	43.28
Yes	41.73	48.02	54.78	56.72
Macrosomia
No	93.25	90.68	85.29	81.89
Yes	6.75	9.32	14.71	18.11
Infant shoulder dystocia
No	99.31	99.04	98.52	98.1
Yes	0.69	0.96	1.48	1.9
Exclusive breastfeeding at delivery
No	71.51	74.94	77.63	79.01
Yes	28.49	25.06	22.37	20.99

Mean time spent in each A1c state

Table 2 displays the unadjusted mean sojourn times a person spent on average following gestational diabetes before transitioning to another A1c state. The average duration spent in a nondiabetes state before transitioning into another state was 2.1 years (95% CI, 2.0-2.1). Those with diabetes following gestational diabetes spent longer in an uncontrolled diabetes state (2.3 years; 95% CI, 2.2-2.4) than a controlled state (1.9 years; 95% CI, 1.8-1.9).

Table 2.

Average sojourn times (years) and 95% CI for average woman with GDM at baseline (n = 34 171).

Diabetes stage	Overall   (N = 34 171)
No diabetes	2.1 (2.0-2.1)
Prediabetes	2.2 (2.2-2.3)
Controlled diabetes	1.9 (1.8-1.9)
Uncontrolled diabetes	2.3 (2.2-2.4)

No diabetes (A1c < 5.7%), prediabetes (5.7 ≤ A1c < 6.5%), controlled diabetes (6.5 ≤ A1c < 7%), and uncontrolled diabetes (A1c ≥ 7%). Average sojourn time refers to average time spent in a given state (in years) before transitioning to another stage.

Adjusted Markov model: Transitions in A1c states.

The probability transition matrix for the adjusted Markov model is shown in Figure 1; that for the unadjusted Markov model is presented in Appendix 3, Table S4. In adjusted models, of participants who did not currently have diabetes following gestational diabetes, most were likely to remain (0.62; 95% CI, 0.62-0.63) in a diabetes-free state the next year (Appendix 3, Table S4). Of those without current diabetes, a modest proportion (0.15; 95% CI, 0.14-0.15) was likely to advance to prediabetes at the next annual assessment. Among those with current prediabetes, a similar proportion was retained in the same status in the next year (0.66; 95% CI, 0.65-0.66). Those with current prediabetes had a moderate probability of reverting back to normoglycemia (0.16; 95% CI, 0.15-0.16) and a low likelihood of progressing to controlled (0.04; 95% CI, 0.03-0.04) or uncontrolled diabetes (0.01; 95% CI, 0.01-0.01) in the next year. In contrast, nearly 1 in 5 of those with a controlled diabetes state were likely to worsen (0.17; 95% CI, 0.16-0.18). The majority of those with controlled (0.61; 95% CI, 0.60, 0.62) or uncontrolled diabetes (0.66; 95% CI, 0.65, 0.67) remained in the same disease status in the next year. Those in an uncontrolled state (0.17; 95% CI, 0.16-0.18) were likely to regain glycemic control. In sensitivity analyses restricted to nulliparous women or that selected the highest A1c value if multiple occurred in a given time interval (Appendix 3, Tables S5-S6), results did not substantially differ.

Figure 1.

Estimated transition probabilities among GDM women, adjusted models (n = 34 171). Adjusted for age group, race/ethnicity, insurance status, educational attainment, and nativity.

Influence of sociodemographic characteristics on A1c transitions

We examined differences in the probability of transitioning between A1c states by baseline sociodemographic factors (Figure 2). Mutually adjusting for all factors, Asian, Black, and Hispanic race/ethnicity (most notably Asian vs White: adjusted HR [aHR] = 1.62; 95% CI, 1.44-.82), foreign born (aHR, 1.16; 95% CI, 1.06-1.28), aged 35 or older (vs 10-24 years) (aHR = 1.31; 95% CI, 1.13-1.52), and a lower educational attainment (aHR = 1.21; 95% CI, 1.10-1.32) were factors associated with a greater likelihood of moving from no diabetes to prediabetes within a 1-year interval. Black and Hispanic individuals had a higher likelihood of a transition from prediabetes to either controlled or uncontrolled diabetes. Among Black participants, for example, the transition hazard ratios for moving from prediabetes to controlled or uncontrolled diabetes were 1.51 (95% CI, 1.24-1.85) and 2.32 (95% CI, 1.21-4.47) higher than among White participants, respectively. In contrast, having Medicaid relative to private insurance and being foreign born relative to US born generally demonstrated a protective influence. For example, Medicaid vs private insurance was associated with a lower likelihood of transitioning from prediabetes to controlled (aHR, 0.77; 95% CI, 0.67-0.88) or uncontrolled (aHR, 0.61; 95% CI, 0.38-0.99) diabetes. Unadjusted and adjusted hazard ratios for the influence of sociodemographic characteristics on A1c stages are presented in Appendix S3, Table S7.

Figure 2.

Forest plot displaying adjusted hazard ratios (95% CI) for the influence of sociodemographic characteristics on transition intensities between diabetes disease states following gestational diabetes (n = 34 171). Full results displayed in Table S7. Dots correspond to adjusted hazard ratio for transition intensities between disease states (presented on log scale) with corresponding 95% CIs (bars). Estimates shown to the left of a null association (gray bar at zero) indicate reduced likelihood of the transition, while indicators to the right of the gray bar indicate characteristics likely to facilitate a transition. Predictors examined individually in models that also control for age group, nativity, race/ethnicity, educational attainment, and insurance status.

Influence of perinatal characteristics on A1c transitions

The influence of preconception health status, pregnancy complications, and adverse birth outcomes on the likelihood of diabetes progression following gestational diabetes is presented in Figure 3 and Appendix 3, Table S8, adjusting for sociodemographic factors. We find that prepregnancy obesity (aHR = 1.48; 95% CI, 1.35-1.62), macrosomia (aHR = 1.19; 95% CI, 1.03-1.36), and, to a lesser extent, chronic hypertension (aHR = 1.13; 95% CI, 0.95-1.35) were associated with a higher likelihood of moving from no diabetes to prediabetes and a lower likelihood of return to normoglycemia (prepregnancy obesity aHR = 0.83; 95% CI, 0.76-0.91; macrosomia aHR = 0.87; 95% CI, 0.75-1.0; prepregnancy hypertension aHR = 0.72; 95% CI, 0.59-0.88). All perinatal characteristics were associated with an increased likelihood of advancing from prediabetes to diabetes (controlled or uncontrolled), with the exception of exclusive breastfeeding at hospital discharge, which had a null association. Perinatal clinical factors also generally were associated with a reduced likelihood of improving from uncontrolled to controlled diabetes, while exclusive breastfeeding showed a marginally significant protective association (aHR = 1.12; 95% CI, 0.98-1.28).

Figure 3.

Forest plot displaying adjusted hazard ratios (95% CI) for the influence of prepregnancy and perinatal clinical characteristics on transition intensities between diabetes disease states following gestational diabetes (n = 34 171). Full results displayed in Table S7. Dots correspond to adjusted hazard ratio for transition intensities between disease states (presented on log scale) with corresponding 95% CIs (bars). Estimates shown to the left of a null association (gray bar at zero) indicate reduced likelihood of the transition, while indicators to the right of the gray bar indicate characteristics likely to facilitate a transition. Predictors examined individually in models also control for age group, nativity, race/ethnicity, educational attainment, and insurance status.

Discussion

We developed a multistate model to characterize state-to-state transitions in hemoglobin A1c levels following gestational diabetes using a population-based cohort followed for up to 9 years. We find that the process that leads to diabetes following gestational diabetes is incremental, without large jumps in A1c levels at 1-year intervals. We find that Black and Hispanic race/ethnicity was generally associated with a higher likelihood of adverse progression and lower likelihood for improvement, although not all transitions were statistically significant. The presence of perinatal comorbidity along with gestational diabetes was also associated with a greater probability of adverse progression in A1c levels and a reduced likelihood of return to normoglycemia.

Our results illustrate the incremental yet intractable progression of A1c levels following gestational diabetes. At 1-year intervals, most individuals (over 6 in 10) had the highest likelihood of remaining in the same A1c state as before. Large jumps in status were rare: moving from prediabetes to uncontrolled diabetes in a 1-year interval, for example, had a 1% probability. Similar jumps in status (eg, from no diabetes to a diabetes state) were not modeled due to data sparseness (Appendix S1, Table S2). While the stability of transition probabilities is favorable for those without current diabetes, it illustrates the relatively intransigent nature of diabetes among those in a more advanced disease stage. For example, those with prediabetes were more likely to return to normoglycemia (16%) than advance to diabetes (5%). However, patients in an uncontrolled diabetes state were more likely to remain in the same status (66%) than to improve (17%). Further, estimated sojourn times show a typical person spent longer in a high A1c state before transitioning to any other state. Those who never had diabetes spent longer in a prediabetes than a no-diabetes state (2.2 years [95% CI, 2.2-2.3] vs 2.1 years [95% CI, 2.0-2.1]). Those who had reached diabetes spent longer in an uncontrolled state (2.3 years; 95% CI, 2.2-2.4) than a controlled diabetes state (1.9 years; 95% CI, 1.8-1.9). Considering prior evidence of worse outcomes with a longer time in uncontrolled diabetes status,⁴⁵ the high retention and longer duration in high A1c states underscore the need for early and intensive prevention and treatment efforts following gestational diabetes to promote glycemic control early in the disease course.

Our results also highlight that the presence of cardiometabolic complications in pregnancy and adverse birth outcomes can inform postpartum risk stratification strategies. We found nearly all perinatal morbidities or adverse outcomes were associated with a greater likelihood of transitioning from prediabetes to diabetes and a lower likelihood of attaining glycemic control among those diagnosed. Exceptions included gestational weight gain and shoulder dystocia (which had a nonsignificant and negative influence, depending on the transition), as well as exclusive breastfeeding before discharge (a nonsignificant or positive influence depending on transition). Notably, exclusive breastfeeding was marginally associated with an increased likelihood of improving to a less severe disease stage, although not influential at all transition points. While the breastfeeding indicator available in our data does not reflect long-term practices and may to some degree reflect hospital practice, the protective effect of breastfeeding is supported by meta-analysis results, which show that breastfeeding following gestational diabetes lessens the likelihood of diabetes.¹² While further study is warranted, results add to prior findings by suggesting breastfeeding is associated with protective changes in A1c at multiple stages, including the transition from normoglycemia to prediabetes.

Results also demonstrate the importance of preconception health on diabetes progression. We found that prepregnancy obesity and chronic hypertension were associated with a larger and more consistent influence on diabetes progression than related disorders that emerged in pregnancy (gestational weight gain and hypertensive disorders of pregnancy). Our findings extend prior evidence of the association between prepregnancy BMI and chronic hypertension with later-life diabetes risk^17,46-48 by illustrating these factors are associated with adverse changes in A1c levels across incremental stages. We found that prepregnancy risk factors, for example, were predictive at early stages of disease progression (from no diabetes to prediabetes, for example), whereas pregnancy-onset conditions were not. The presence of preconception cardiometabolic dysregulation may signal preexisting risk, leading to more rapid advancement through disease stages than those with a pregnancy-onset complication.^2,47,49,50 For example, many pregnancy complications are preceded by subclinical vascular and metabolic dysfunction, indicating latent dysregulation.⁵¹ Together, results suggest preconception health can inform targeted postpartum A1c monitoring efforts.

Prior research has found a higher burden of diabetes among Black and Hispanic people following gestational diabetes.⁹ We add to this by showing that Black and Hispanic race/ethnicity was associated with a higher likelihood of increasing A1c levels and a generally lower likelihood of improvement. One potential explanation for this finding is that Black and Hispanic individuals in the United States disproportionately experience the effects of structural racism, which contributes to neighborhood disadvantage and psychosocial stress and impedes access to care.^52,53 The cumulative effects of chronic disadvantage and discrimination may, for example, contribute to β damage, which is key to the development of type 2 diabetes following gestational diabetes.^15,54 The stronger transition to an elevated A1c state observed in this study may contribute to later-life disparities in cardiovascular disease, which disproportionately affect Black and Hispanic groups.^55-58 Notably, our results also show that having Medicaid relative to private insurance was associated with beneficial influences on A1c trajectories. Other sociodemographic factors had mixed associations depending on A1c state. Being foreign born and having lower education, for example, were associated with an increased probability of transitioning to prediabetes from no diabetes but showed a protective association in later phases of disease. National study findings among Latinos (2013-2018) also depict a pattern of higher prediabetes among foreign-born relative to US-born adults.⁵⁹ In contrast to our findings on A1c transitions, Black, Latino, and Asian immigrants bear a higher diabetes burden than their US-born counterparts.⁸ Taken together, results illustrate heterogeneity in A1c transitions by sociodemographic profiles.

The strengths of this study include use of a diverse population-based cohort with longitudinal A1c data for up to 9 years (the APPLE NYC Cohort). Where possible, we defined study exposures using a combination of birth record or via hospital diagnosis codes at delivery.²⁵ However, all study covariates are subject to misclassification. A1c levels are an incomplete proxy for medical diagnosis. Information on postpartum medication or lifestyle interventions (eg, duration or exclusivity of breastfeeding, postpartum weight loss) was not available, which may in part explain observed disparities in disease progression.⁶⁰ We note that the form of the multistate model represents the underlying progression of A1c values over time, as opposed to the observed progression. Exact transition times are unknown; rather, they are interval-censored (ie, snapshots of the underlying process). Markov models rely on the memoryless property assumption, which anticipates the next likely disease phase based on an individual’s current health state. Both empirical results provided in the supplement as well as the fact that providers often have incomplete medical record history for patients support the reasonableness of this assumption.

Several sensitivity analyses conducted using APPLE Cohort data support the robustness of our results. While we present results from the vantage point of a given gestational diabetes pregnancy, analyses restricted to women who have never been pregnant demonstrate that ruling out the potential for a previous gestational diabetes pregnancy has little influence on study results. Previous analyses from this cohort have also shown that multiple gestational diabetes pregnancies did not appreciably change diabetes risk.⁶ Additionally, we found that 44% of those with gestational diabetes at index pregnancy individuals did not receive repeated postpartum A1c testing (at least 2 tests). Reasons for loss of follow-up could include screening bias, relocation from NYC, and receiving care outside of NYC, among other factors. Comparison of sociodemographic characteristics among those who receive and do not receive repeated postpartum A1c testing (Table S1) suggests that our results may represent individuals with higher diabetes risk than those who did not receive postpartum screening. Previous sensitivity analyses of the APPLE cohort have demonstrated the robustness of cohort data against potential differential screening practices and loss to follow-up under a variety of scenarios.^6,7

Conclusions

We characterized longitudinal transitions in hemoglobin A1c levels following gestational diabetes up to 9 years postpartum as well as predictors of disease transitions. Findings support growing evidence that indicates the utility of sociodemographic and perinatal characteristics as useful markers for identifying women at high risk of adverse postpartum transitions in hemoglobin A1c. In particular, transitions among persons of color were marked by a stronger likelihood of moving to a higher A1c state and a reduced likelihood of improvement. Those with prepregnancy obesity and hypertension may also benefit from enhanced postpartum monitoring. Better understanding of the natural progression of postpartum A1c levels following gestational diabetes highlights the importance of the disruption of diabetes progression at each of its incremental stages.

Supplementary material

Supplementary material is available at the American Journal of Epidemiology online.

Funding

Supported by National Institutes of Health grants R01DK134725 and R21DK122266. The National Institutes of Health had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.

Conflict of interest

The authors report no potential conflicts of interest relevant to this article.

Data availability

The data that support the findings of this study are available on reasonable request from the study principal investigator (T.J.) and are subject to approval by the NYC Department of Health and Mental Hygiene. Data are not publicly available due to the sensitive nature of the data and to limit potential risks of reidentification.