Low Carbohydrate–Diet Scores and Long-term Risk of Type 2 Diabetes Among Women With a History of Gestational Diabetes Mellitus: A Prospective Cohort Study

Type 2 diabetes mellitus (T2DM) has become a worldwide epidemic, underscoring the importance of preventing T2DM as a public health priority (1). Women with a history of gestational diabetes mellitus (GDM), a common pregnancy complication, represent a group at high risk of T2DM (2). Specifically, women with GDM have sevenfold increased risk of developing T2DM compared with those with a normoglycemic pregnancy (3). Although GDM is detected during pregnancy, most women with GDM probably have reduced insulin secretion and/or chronic insulin resistance before pregnancy, and the glucose intolerance may deteriorate through their life span (2). The roles of dietary and lifestyle factors in modifying the natural course of the progression from GDM to overt T2DM remain to be elucidated in such a high-risk population. In previous studies, we found that higher diet quality and a physically active lifestyle are associated with lower risk of T2DM among women with prior GDM (4,5).

Low-carbohydrate diets (LCDs) are among the most popular diets that are claimed to promote weight loss (6). These diets are appealing because of not having to count calories or compromise the consumption of many palatable foods. However, LCDs are at risk for being nutritionally inadequate or imbalanced because they enforce restriction of food choices (7). These diets limit consumption of some healthful dietary components, such as whole grains, dietary fiber, fruit, and vegetables, and they can be high in animal fat and red meat, which have been associated with an increased risk of T2DM (8). Therefore, long-term adherence to LCDs, particularly those mainly based on animal-source foods, may have unfavorable effects on glucose homeostasis and increase the risk of T2DM. A previous study found a dietary pattern score representing moderately low carbohydrate, high animal protein, and high animal fat intakes was positively associated with T2DM risk (9).

Women with GDM are advised to limit carbohydrate intake during and after pregnancy as a key component of medical nutrition therapy. For instance, the Endocrine Society Clinical Practice Guideline suggests that women with GDM should limit carbohydrate intake to 35–45% of total calories (10). LCDs may improve short-term glycemic control in patients with GDM (11,12), but the long-term effect on the progression from GDM to T2DM is unknown. In this study, we examined the long-term risk of T2DM in association with three low-carbohydrate dietary patterns among women with a history of GDM, a population at high risk of T2DM.

The study population is composed of women with a history of GDM in the Nurses’ Health Study II (NHSII), as part of the ongoing Diabetes & Women’s Health (DWH) study (13), which aims to identify determinants of the progression from GDM to T2DM. The NHSII, established in 1989, is an ongoing prospective cohort study of 116,430 female nurses aged 24–44 years at study initiation (14). The participants receive a biennial questionnaire to update information on health-related behaviors and disease outcomes. Follow-up for each questionnaire cycle is greater than 90%. This study was approved by the Partners Human Research Committee (Boston, MA), with participants’ consent implied by the return of the questionnaires.

In this cohort, dietary information was collected first in 1991 and every 4 years thereafter using a semiquantitative food-frequency questionnaire (FFQ). Participants were eligible for inclusion in this analysis if they reported GDM before 1991 or incident GDM through the 2001 questionnaire. GDM was last captured on the 2001 questionnaire because most of the NHSII participants passed reproductive age. In a prior validation study among a subgroup of this cohort, 94% of GDM self-reports were confirmed by medical records (14). In a random sample of parous women without GDM, 83% reported a glucose screening test during pregnancy, and 100% reported frequent prenatal urine screenings, indicating a high level of GDM surveillance in this cohort (14). Participants were excluded from this analysis if they reported chronic disease (T2DM, cardiovascular disease, cancer) before their GDM pregnancy or before the return of their first post-GDM FFQ, had a multiple-birth pregnancy, or did not return any post-GDM FFQ.

Every 4 years since 1991, the participants reported their usual intake (frequency ranging from never to >6 times/day) during the past year of a standard portion of each item in the FFQ. The reproducibility and validity of the FFQ has been extensively documented (15–17). We computed intakes of individual nutrients by multiplying the frequency of consumption of each food by the nutrient content of the specified portion based on food composition data from U.S. Department of Agriculture and data from manufacturers. Intakes of carbohydrate, fat, and protein were expressed as nutrient densities (i.e., percent of energy), and intakes of other nutrients were energy adjusted using the residual method (18). A previous validation study comparing energy-adjusted macronutrient intake assessed by the FFQ with four 1-week diet records found the Pearson correlation coefficients were 0.61 for total carbohydrates, 0.52 for total protein, and 0.54 for total fat (16).

To represent adherence to various low-carbohydrate dietary patterns, we calculated three LCD scores (i.e., overall LCD score, animal LCD score, and vegetable LCD score) for each participant, as previously described (19,20). Briefly, we divided study participants into 11 strata according to each of fat, protein, and carbohydrate intake, expressed as percentages of energy. We assigned them 0–10 points for increasing intake of total fat, 0–10 points for increasing intake of total protein, and inversely, 10–0 points for increasing intake of carbohydrates. Then we summed points for the three macronutrients to create an overall LCD score, which ranged from 0 to 30. Similarly, we also created an animal LCD score to represent preferential substitution of carbohydrates with fat and protein from animal sources and a vegetable LCD score to represent preferential substitution of carbohydrates with fat and protein from vegetable sources. The animal LCD score was based on intakes of carbohydrate, animal protein, and animal fat, and the vegetable LCD score was based on intakes of carbohydrate, vegetable protein, and vegetable fat. A higher score reflects a higher intake of fat and protein and a lower intake of carbohydrate, indicating closer adherence to a low-carbohydrate dietary pattern.

Information on age, weight, race/ethnicity, family history of diabetes, smoking status, age at first birth, use of oral contraceptives, and menopausal status has been collected on biennial questionnaires. Parity was defined as the number of pregnancies lasting greater than 6 months. Self-reported weight was highly correlated with a measured weight (r = 0.97) in a previous validation study (21). BMI was computed as weight in kilograms divided by height in meters squared. Total physical activity was ascertained by frequency of engaging in common recreational activities, from which MET h/week were derived. In a previous validation study in this cohort, the correlation coefficients of physical activity as reported in 1-week recalls and diaries with that reported on the questionnaires were 0.79 and 0.62, respectively (22).

Participants reporting physician-diagnosed T2DM on each biennial questionnaire were mailed a supplemental questionnaire regarding symptoms, diagnostic tests, and hypoglycemic therapy to confirm self-reported diagnoses. Confirmed T2DM was defined according to the American Diabetes Association criteria (23): 1) one or more classic symptoms (excessive thirst, polyuria, unintentional weight loss, hunger) plus elevated glucose levels (fasting plasma glucose concentration ≥7.0 mmol/L or random plasma glucose ≥11.1 mmol/L), 2) no symptoms reported but two or more elevated plasma glucose concentrations on more than one occasion (fasting ≥7.0 mmol/L, random ≥11.1 mmol/L, 2-h oral glucose tolerance test ≥11.1 mmol/L), or 3) treatment with insulin or oral hypoglycemic agent. Before 1998, fasting plasma glucose ≥7.8 mmol/L was used instead of ≥7.0 mmol/L for the diagnosis of diabetes according to the criteria of National Diabetes Data Group (24). A subgroup validation study conducted in a similar cohort of U.S. female nurses reported a high accuracy rate of 98% by comparing our classification against medical records (25).

In this analysis, we defined baseline as the questionnaire period in which participants first reported a GDM pregnancy. The baseline characteristics of the cohort were age adjusted according to 5-year age-groups by direct standardization to the entire cohort. Comparisons among quintiles of LCD scores were performed using ANOVA for continuous variables and the χ² test for categorical variables.

We computed follow-up time from the date of GDM diagnosis to the date of T2DM diagnosis, death, or the return of the 2011 questionnaire, whichever came first. Updating of exposure status ceased if a participant reported incident chronic diseases (cardiovascular disease, cancer). We carried forward missing exposure data from the most recent questionnaire for which data were captured. Exposure data were missing for less than 7% of the participants. To represent long-term habitual diet after GDM and reduce measurement error (26), we calculated a cumulative average LCD score based on the information from each post-GDM FFQ during the study period.

We used Cox proportional hazards models to estimate the hazard ratios (HRs) and 95% CIs of T2DM risk in relation to quintiles of LCD scores. The covariates adjusted in the models included age (months), parity (1, 2, 3, ≥4), age at first birth (12–24, 25–29, ≥30 years), race/ethnicity (Caucasian, African American, Hispanic, Asian, other), family history of diabetes (yes or no), oral contraceptive use (current, former, never), menopausal status (premenopausal or postmenopausal), cigarette smoking (current, former, never), alcohol intake (0, 0.1–4.9, 5.0–14.9, ≥15.0 g/day), total energy intake (quintiles), and glycemic index (quintiles). All of the time-varying covariates were updated during the follow-up and were modeled as time-varying variables in the models. We included BMI (<23.0, 23.0–24.9, 25.0–26.9, 27.0–29.9, 30.0–34.9, 35.0–39.9, ≥40.0 kg/m²) in the model separately because BMI may be an intermediate variable between the LCD scores and incident T2DM. We also performed mediation analyses to estimate the proportion of the association between the LCD scores and incident T2DM that is explained by BMI (modeled continuously) (27), using a publicly available SAS macro (28).

To evaluate potential effect modification, we performed stratified analyses according to age (<40 or ≥40 years), family history of diabetes (yes or no), obesity (BMI <30 or ≥30 kg/m²), and time since the first GDM pregnancy (>10 or ≤10 years). We conducted interaction tests via multiplicative interaction terms in the multivariable models. All statistical analyses were performed with SAS 9.3 software (SAS Institute Inc., Cary, NC). P < 0.05 was considered statistically significant.

We documented 722 incident T2DM cases from 4,502 participants with a history of GDM (3,243 prevalent cases of GDM in 1991 and 1,259 incident GDM cases developed between 1991 and 2001), contributing 68,897 person-years of observation. The average number of pregnancies with GDM was 2.5 among women who had a diagnosis of T2DM and 2.0 among women who did not have a diagnosis of T2DM. At baseline, women with a higher overall LCD score were more likely to be white, currently smoking, and less physically active; to have a higher BMI and family history of diabetes; and to consume more cholesterol and heme iron but less total calories and dietary fiber (Table 1). By food group categories, they consumed more red meat and less fruits, vegetables, and whole grains. We observed similar results for the animal LCD score (Supplementary Table 1). In contrast, participants with a higher vegetable LCD score consumed more nuts, vegetables, and whole grains but less calcium than those with a lower score (Supplementary Table 2). Each of these three LCD scores was inversely associated with dietary glycemic index and glycemic load. There was a very high correlation between the overall LCD score and the animal LCD score (r = 0.93, P < 0.001), and their correlations with vegetable LCD score were also significant (r = 0.49 for correlation between overall and vegetable LCD scores, P < 0.001; r = 0.20 for correlation between animal and vegetable LCD scores, P < 0.001). The animal LCD score had the most statistical variation (standard deviation [SD] 8.1), followed by overall LCD score (SD 7.5) and vegetable LCD score (SD 5.6), although they all had a mean score of 15 in this study population.

Overall and animal LCD scores were positively and significantly associated with risk of T2DM among women with a history of GDM, whereas the vegetable LCD score was not significantly associated with the risk (Table 2). The multivariable-adjusted HRs (95% CIs) of T2DM comparing the highest with lowest quintiles were 2.13 (1.65–2.76) for overall LCD score (P < 0.001 for trend), 2.18 (1.68–2.83) for animal LCD score (P < 0.001 for trend), and 1.29 (1.00–1.67) for the vegetable LCD score (P = 0.14 for trend). The significant associations of overall and animal LCD scores with T2DM risk were attenuated but remained significant after additional adjustment for updated BMI, with the HRs (95% CIs) comparing the highest versus the lowest quintiles of 1.36 (1.04–1.78; P = 0.003 for trend) for overall LCD score and 1.40 (1.06–1.84; P = 0.004 for trend) for the animal LCD score. Mediation analyses estimated that updated BMI explained 66% (95% CI 42–89; P < 0.001) and 64% (95% CI 42–86; P < 0.001) of the total effects of the overall LCD score and the animal LCD score on T2DM risk, respectively. In addition, the associations of the overall LCD score and the animal LCD score with T2DM risk were not significantly altered by age, family history of diabetes, obesity status, and time since GDM pregnancy (all P > 0.10 for interaction). Similar results were observed in sensitivity analyses by skipping questionnaire cycles in which FFQs were not returned or carrying forward missing exposure data from the participant’s cumulative average intake of all past post-GDM FFQs instead of carrying forward missing exposure data from the most recent post-GDM FFQ available as in the current analysis.

In this prospective cohort study with up to 20 years of follow-up, we observed that a dietary score representing a low-carbohydrate, high–animal protein, and high–animal fat dietary pattern was significantly and positively associated with T2DM risk among women with a history of GDM. These associations were partly explained by BMI. By contrast, a dietary score representing a low-carbohydrate, high–vegetable protein, and high–vegetable fat dietary pattern was not associated with the risk of developing T2DM. The observed associations were not significantly modified by age, family history of diabetes, smoking, obesity status, or time since the first GDM pregnancy.

Previous studies examining the effects of carbohydrate restriction on glucose metabolism have yielded conflicting results. A nonrandomized study reported that carbohydrate restriction in patients with diet-controlled GDM resulted in improved glycemic control and less need for insulin therapy after 6 weeks of diet therapy (11). However, LCD treatment did not significantly reduce the need for insulin therapy among women with GDM in a recent randomized controlled trial with 4 months of follow-up (12). In addition, several randomized controlled trials have shown that LCDs seem to yield a significant reduction in blood glucose concentrations at 6 months after the intervention but that the reduction becomes nonsignificant at 12 months (29–33).

Women who follow low-carbohydrate dietary patterns consume less carbohydrate and more fat and protein that are naturally needed to compensate energy requirements. A previous study (34) showed a null association of quantity of carbohydrate intake with T2DM risk. The observed positive associations of overall and animal LCD scores with risk of T2DM in this study may be at least partly owing to some detrimental effects of a high content of animal fat and/or animal protein. Previous studies in animal models and in humans show that higher intake of dietary animal fat can result in impaired glucose tolerance and T2DM (35,36). Higher intake of animal protein was associated with an increased risk of T2DM (37–39). Moreover, a meal rich in animal protein, compared with a meal rich in vegetable protein, results in higher plasma concentrations of branched-chain amino acids (40), which have been recently associated with insulin resistance and the development of T2DM in several metabolomics studies (41,42).

Several food groups, including red meat, vegetables, fruits, and whole grains, may contribute to the variability of the LCD scores and account for the observed associations. Red meat is a major dietary source of animal protein and animal fat in the Western diet. In this study, women with a higher overall or animal LCD score had a greater consumption of red meat than those with a lower score. Higher consumption of red meat has been previously associated with an increased risk of T2DM (43). On the other hand, women with a higher overall or animal LCD score, compared with those with a lower score, had a lower consumption of fruits, vegetables, and whole grains, which have been inversely associated with the risk of T2DM in previous studies (44,45).

Our study has several strengths, including the prospective cohort design that establishes the temporal direction of the associations, the large sample size, the long-term follow-up, the high response rates (> 90%) of each questionnaire cycle in both the entire cohort and the subcohort members included in this study, and the detailed prospective dietary assessments with the extensively validated FFQs (15–17). All the study participants were registered nurses, reducing the potential confounding by educational attainment or differential access to health care.

We acknowledge that there are several limitations. First, misclassification of dietary intakes of carbohydrate, fat, and protein is possible. However, the misclassification would be nondifferential due to the prospective nature of the dietary assessment with regard to incident T2DM; therefore, our findings may underestimate the true associations. Furthermore, the use of cumulative averages of dietary intakes for participants with longitudinal FFQs reduces random error.

Second, screening bias may exist if more health-conscious women regularly visit a physician, thus increasing their chance of receiving a medical diagnosis. However, in our sensitivity analyses, similar results were seen when we restricted cases to symptomatic T2DM, minimizing concerns for this bias.

Third, our study population consisted mostly of Caucasian American women. Thus, the direct generalization of our findings to other populations whose major food sources of macronutrients are different may be limited. Among our participants, Asian women had lower LCD scores (i.e., higher carbohydrate intake) compared with Caucasian, African American, and Hispanic women. In contrast to findings among U.S. men (9) and women (46), a recent study among the Japanese population, in which white rice consumption is high, found a LCD score was associated with lower risk of T2DM (47). The association between LCD scores and risk of T2DM across different race/ethnic groups warrants further evaluation.

In conclusion, among women with a history of GDM, a low-carbohydrate dietary pattern, particularly with high protein and fat intake mainly from animal-source foods, is associated with higher T2DM risk, whereas a low-carbohydrate dietary pattern with high protein and fat from plant-source foods is not associated with risk of T2DM. Women with a history of GDM who follow a low-carbohydrate dietary pattern may consider consuming plant sources rather than animal sources of protein and fat to minimize their future risk of T2DM.

Acknowledgments. The authors thank Thomas L. Halton (Department of Nutrition, Harvard T.H. Chan School of Public Health) for devising the LCD scores.

Funding. This study was supported by the Intramural Research Program of the Eunice Kennedy Shriver National Institute of Child Health and Human Development, National Institutes of Health (contract no. HHSN275201000020C). The Nurses’ Health Study II was funded by research grants DK-58845, CA-50385, P30-DK-46200, and UM1-CA-176726 from the National Institutes of Health. D.K.T. was supported by a mentored fellowship from the American Diabetes Association (no. 7-12-MN-34).

Duality of Interest. No potential conflicts of interest relevant to this work were reported.

Author Contributions. W.B. contributed to the design and analysis of the study and wrote the manuscript. S.L. conducted technical review, interpreted the results, and reviewed and edited the manuscript. J.E.C., D.K.T., Y.Z., and F.B.H. interpreted the results and reviewed and edited the manuscript. C.Z. contributed to the design and analysis of the study and reviewed and edited the manuscript. W.B. and C.Z. are the guarantors of this work and, as such, had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis.

Prior Presentation. Parts of this study were presented as an abstract at the 75th Scientific Sessions of the American Diabetes Association, Boston, MA, 5–9 June 2015.