Women with gestational diabetes mellitus (GDM) have a 7-10-fold higher risk of incident diabetes (1) and an increased risk for future cardiovascular disease (2, 3). However, the population of GDM, just as observed in the population of subjects with diabetes, is not homogenous (4, 5). Early identification of women with GDM at the highest risk of future diabetes is crucial for prevention and long-term follow up (6).

Diabetes heterogeneity has recently been illustrated by a new classification system based on six variables (age, BMI, HbA1c, an estimate of the homoeostasis model assessment of β-cell function (HOMA-B), insulin resistance (HOMA-IR) and the presence or absence of islet autoantibodies) with distinct risks for complications (7, 8). The potential clustering of these variables and its impact on future diabetes risk has not been investigated in women with GDM.

Although GDM is generally characterized by increased insulin resistance and reduced insulin secretion (9), considerable heterogeneity exists (4, 5). Compared to women with normal glucose tolerance, half of women with GDM have increased insulin resistance and a third have reduced insulin secretion (4). While a previous study found that women with reduced insulin secretion did not differ from their normal glucose tolerant counterparts regarding BMI, fasting glucose in pregnancy, and the risk of adverse obstetric and neonatal outcomes, those with higher insulin resistance had a higher BMI, fasting glucose, and a higher risk of adverse pregnancy outcomes (4, 10).

Higher maternal age, pre-pregnancy weight and BMI, early diagnosis of GDM, the need for insulin treatment during pregnancy, glucose level and HbA1c in the third trimester have been identified as predictors of future diabetes or prediabetes in women with GDM (11–13). Other predictors of future glucose intolerance (GI) include a family history of type 2 diabetes, high-risk ethnicity, and previous GDM (14, 15). Although eating behaviour (intuitive eating) or dietary intake such as low-fat and low protein diet can predict reduced future diabetes risk after GDM (16, 17), the independent role of nutritional intake or eating behavior among women with GDM during pregnancy on future glucose intolerance or diabetes has not been previously investigated. Only few studies have investigated predictors of GI in the postpartum and this has been shown mainly in Asian populations (18). Previous studies have focused on predictors for all women together. However, predictors of GI might vary according to GDM subtypes (4, 10, 19). To account for the heterogeneity of GDM, more precise insight into different predictors according to GDM clusters or BMI-categories can help to better understand different risk factors within each subgroup.

We studied an exhaustive list of predictors (both in pregnancy and in the early postpartum) of GI in all women with GDM together and in GDM subgroups. Subgroups were investigated according to (a) a data-driven cluster analysis based on age, BMI, HOMA-IR and HOMA-B indices (8); and for clinical ease according to (b) normal weight (NW) vs overweight or obesity (OW-OB). This was done to identify the characteristics of the clusters and their metabolic risk profile and to propose a tailored approach for screening and treatment.

This study used an adaption of the newly proposed diabetes-cluster classification to classify women with GDM based on age, BMI, HOMA-IR and HOMA-B. We compared this classification to a clinical classification according to BMI-categories. During pregnancy and in the postpartum, we identified both an “insulin-resistant” and an “insulin-deficient” cluster of women with GDM. The former had a 2.9-fold (for pregnancy) to 3.5-fold (for the early postpartum) increased risk of GI at 1-year postpartum. Tailored predictors in pregnancy for GI at 1-year postpartum were a history of GDM and fasting glucose for the “insulin-deficient” cluster, as well as for women in the NW BMI-category. For the “insulin-resistant” cluster or women with OW/OB, this was HOMA-IR. In the postpartum, we confirmed the utility of 1-h glucose during early postpartum as predictive for future GI especially when a more complex measure, ISSI-2 (or insulin sensitivity-adjusted insulin secretion), was removed from the analysis.

Overall, our data-driven analyses point towards the existence of two clusters with different risks of future GI in women with GDM. For clinical ease, BMI-category, particularly NW could represent a simple proxy for the “insulin-deficient” cluster, as we found a very high concordance between women with NW and the “insulin-deficient” cluster. In addition, belonging to the “insulin-resistant” cluster (as compared to the “insulin-deficient” one) increased the risk of GI at 1-year postpartum to a similar extent as belonging to the OW/OB category did when compared to NW women. Predictors of future GI also varied according to clusters and BMI-categories and were particularly specific and helpful in pregnancy, which points to the utility of a more personalized approach.

While Ahlqvist et al. classified patients with diabetes into five clusters with different characteristics and risks of diabetic complications outside pregnancy, our analyses in women with GDM identified two clusters. The characteristics of three of their clusters resemble the ones found in our population (severe insulin-deficient diabetes, mild obesity-related diabetes, severe insulin resistant diabetes). In analogy to Ahlqvist et al, glucose control also differed between our “insulin-resistant” cluster. Most women (98%) belonging to “insulin-deficient” cluster during pregnancy were also categorized as “insulin-deficient” in the postpartum. The concordance between the “insulin-resistant” clusters during pregnancy and the postpartum was 67%, i.e. a third of women in the “insulin-resistant” cluster during pregnancy “switched” the “insulin-deficient” cluster in the postpartum. This probably means that a third of women have a particularly pronounced increase in insulin resistance during pregnancy that is transient and reverses in the postpartum. In terms of complications, our “insulin-resistant” clusters had a 2.9-3.5 higher risk of later GI. Data-driven cluster analyses in this population may be useful for future metabolic risk assessment. For clinical feasibility, we also assessed the risk of GI in BMI-categories and could confirm similar results, i.e., we found higher risks in women with OW/OB, compared to NW women. This suggest that BMI, a simple and cheap measure, could serve as a proxy for part of the cluster analysis in this population.

Another study has used cluster analysis in pregnancy to assess the risk of offspring obesity in a general population of healthy pregnant women (34). Based on nine laboratory variables, they identified five clusters, among others an insulin resistant–hyperglycemic cluster that had the highest rates of obesity in the offspring. Previous studies in women with GDM did not perform cluster analysis, but performed subgroups, that were based on insulin resistance (Matsuda index or HOMA-IR) or on more complex insulin secretion measures (Stumvoll 1st phase index) (4, 10, 19, 35). They mainly investigated the relationship between GDM subgroups and pregnancy outcomes (10, 19, 36). Of those focusing on future risk of GI, two showed no differences between the subgroups (10, 35). We are only aware of one study that found differences in the risk of developing GI between subgroups of women with GDM. Their subgroups were based on family history of diabetes, the need for glucose-lowering treatment during pregnancy, and pre-pregnancy BMI (37). However, these studies (10, 35) did not examine tailored predictors of GI within these subgroups.

To further provide a more tailored insight and counseling, we assessed the predictors of GI according to clusters and BMI-categories. Most studies assessed predictors of GI for the entire population, grouping women with NW, OW and OB together (12, 38, 39). Similar to a few other studies (40, 41) we considered women with OW and OB separately from those with NW, but pooled them as one BMI category (OW/OB) and compared them to women with NW. Indeed, there is evidence of similar pattern of altered signaling pathways in the human foetoplacental microvascular and macrovascular endothelium in women with OW and OB, which contrasts with findings in NW women (42). In pregnancy, results showed that fasting glucose and a history of GDM were the most relevant predictors of GI for the “insulin-deficient” cluster, and in women with NW. For the “insulin-resistant” cluster and the OW/OB category it was insulin resistance (HOMA-IR). Thus, predictors of GI in pregnancy were similar within clusters or BMI-categories and could be attributed to the differences in insulin resistance or deficiency. This was additionally confirmed in supplementary analyses for low/high HOMA-IR and HOMA-B subgroups. We also showed that detailed evaluation of eating behavior or nutritional intake in pregnancy were not superior to simple measures. Previous studies have shown that higher vegetables or fruit intake as well as lower energy, protein, and animal fat intake in the postpartum was associated with lower risk of GI (43), but their results were not independent of other clinical and laboratory parameters like glucose control and insulin resistance/sensitivity indices.

In the early postpartum, ISSI-2 was the most important predictor of GI for the “insulin-deficient” cluster. Interestingly, the 1-h glucose level during the oGTT was an important predictor for all women together, and for the NW women. When removing the more complex ISSI-2 from the model, the 1-h glucose predicted GI in women with OW/OB. This is in line with findings of Göbl et al. (29), who showed that 1-h glucose levels during the oGTT in the postpartum were independently associated with a risk for overt diabetes among women with GDM up to 10-years of follow-up. Our results confirm the utility of the 1-h glucose value in the early postpartum as a predictor of future GI.

We propose a novel extension of the new diabetes-classification system to include women with GDM. Other strengths of this study are the use of tailored predictors for future GI based on BMI-categories and clusters, both in pregnancy and in the early postpartum. Compared to other studies, we used a very exhaustive list of clinical, laboratory and more complicated indices of insulin secretion/resistance as well as dietary intake and eating behaviour.

Limitations of our study include a relatively small sample size, particularly in some stratified subgroups, and the fact that we pooled women in the intervention and control group together. However outcomes and effect sizes were not different between groups, and we tested for interactions and always adjusted for group allocation. Additionally, we analyzed women with OW and OB as a single group, which might result in the loss of data. We did this to increase our sample size, as for some measures such as the FFQ or detailed measures of insulin resistance and sensitivity we would not have enough valid measures. A further limitation is the focus on pre-pregnancy BMI and not on body composition when categorizing women. However, we did not have any data about body composition before pregnancy. Furthermore, using body composition would be more cumbersome in a clinical setting. Of note, 80% of women in the OW/OB category based on pre-pregnancy BMI were also attributed to the category of high-fat (based on the fat mass in pregnancy assessed by bioimpedance: low, high fat mass using the median value as cut-off ≤ 30.87 kg) and 85% of women in the NW category based on pre-pregnancy BMI also to the low-fat category. Women were only followed-up until 1 year postpartum. Another limitation is the subjective assessment of dietary intake, using an FFQ. Moreover, several measures such as nutritional intake or the more complicated indices of insulin resistance and secretion were not assessed at all time points and there were some missing data, especially for more complicated indices. The unsupervised rather than supervised approach taken in this study for the clustering analysis was due to the fact that our data was unlabeled and we were interested to describe groups, patterns and relationships. With this approach, there is no use to train and validate our current model in a separate population, but another approach could be used in future studies. The choice of a k-means algorithm was mostly to align our approach with previous studies. We also used a k-medoids analysis and got similar patient assignments to clusters (results not reported in this study). As this cluster classification of women with GDM has been only shown in this study, a replication of our results would be needed to increase its generalizability and external validity.

Overall, we identified two clusters in women with GDM (an “insulin-resistant” and an “insulin-deficient” cluster) with distinct risks of future GI. Predictors varied according to clusters and BMI-categories emphasizing the need for tailored risk assessments.

Amar Arhab, Pascal Bovet, Arnaud Chiolero, Stefano Di Bernardo, Adina Mihaela Epure, Sandrine Estoppey Younes, Leah Gilbert, Justine Gross, Antje Horsch, Stefano Lanzi, Seyda Mayerat, Yvan Mivelaz, Jardena J Puder, Dan Yedu Quansah, Jean-Benoit Rossel, Nicole Sekarski, Umberto Simeoni, Bobby Stuijfzand, Yvan Vial.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by La Commission cantonale d’éthique de la recherche sur l’être humain (CER-VD). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

AL-S: Conceptualization, Data curation, Investigation, Methodology, Software, Visualization, Formal analysis, Writing – original draft. EC: Writing – review & editing. AL: Formal analysis, Writing – review & editing. SS: Writing – review & editing. DYQ: Conceptualization, Data curation, Methodology, Visualization, Supervision, Writing – review & editing. JJP: Conceptualization, Data curation, Methodology, Supervision, Funding acquisition, Project administration, Visualization, Writing – review & editing.