Introduction
Insulin resistance is a typical trait of pregnancy, even when not complicated by any endocrinological, metabolic or cardiovascular disorder [
1‐
6]. Of note, when pregnancy is complicated by gestational diabetes mellitus (GDM), insulin resistance further deteriorates [
7‐
10]. According to international guidelines, GDM can be diagnosed by fasting glycemia or by an oral glucose tolerance test (OGTT) [
11]. With regard to the OGTT, GDM diagnosis requires collecting only the 60 min and the 120 min blood samples following glucose ingestion, and only glycemia needs to be assessed [
11]. Thus, in routine clinical practice, the OGTT includes only two time samples in addition to the fasting sample, and with no measurement of insulin or C-peptide. In fact, more complete OGTT studies are currently limited to clinical trials rather than the clinical routine. However, the information of the simple diagnostic OGTT do not allow assessment of insulin resistance, which requires insulin measurement in addition to glucose and takes advantage from possible collection of more blood samples rather than the 60 and 120 min samples only [
12‐
17]. Nonetheless, the assessment of insulin action may be clinically relevant in GDM and even in pregnant women with overweight, obesity or with signs of dysglycemia (though in the absence of overt GDM), since the degree of insulin resistance can drive specific therapeutic intervention and care [
18,
19]. In addition, though there are still controversies about optimal timing for metabolic assessment in pregnancy [
20,
21], the assessment at early pregnancy (i.e., before 24 weeks of gestation) of the insulin resistance level (and, possibly, other metabolic parameters) may allow prediction of GDM risk and severity of the disease at later stage. Thus, the problem arises on how to assess insulin resistance or sensitivity without the OGTT or even more complicated tests, such as the hyperinsulinemic-euglycemic clamp, which are hardly feasible in the clinical practice (especially in pregnant women).
There is an increasing interest in the triglyceride-glucose index (TyG) [
22], as mirrored by the remarkably growing number of articles dealing with TyG in the recent years. Especially, in the 2019–2021 period the number of pertinent articles has increased each year by about the 50% compared to the previous year, and at mid-2022 the number of articles has already reached that of the whole 2020 (source: PubMed). This is likely due to the relevant properties of TyG, despite its simple formulation, as it only relies on fasting triglycerides and glucose [
22]. The importance of TyG has in fact been summarized in several review studies, which have shown TyG ability to predict risk for cardiovascular diseases, type 2 diabetes and GDM as well [
23‐
32]. In addition, TyG has been proposed in several investigations as surrogate marker of insulin resistance [
24,
33,
34]. TyG as surrogate of insulin resistance appears in fact the original reason for its introduction in the scientific community [
22,
35]. Notably, insulin resistance is a relevant aspect of the metabolic syndrome, and, as such, an important risk factor for cardiovascular diseases or diabetes [
36‐
46]. However, studies on TyG potential and relevance in pregnancy are still relatively scarce [
30,
47‐
55]. Furthermore, studies on TyG performance as surrogate marker of insulin resistance in pregnancy are even rarer [
47,
51], and limited to the assessment of insulin resistance at fasting rather than in dynamic conditions (i.e., during a glucose or meal challenge), which allow more reliable assessment of the actual degree of whole body insulin resistance. On the other hand, in a previous study from our research group we found that in pregnant women fasting triglycerides are associated with insulin resistance derived from an OGTT [
56]. However, we did not investigate the relationships between OGTT-derived insulin resistance and TyG.
Thus, one aim of the present study was to explore OGTT-derived insulin resistance and TyG relationships in pregnancy. To our knowledge, in pregnancy no previous study analyzed the performance of TyG as marker of insulin resistance derived by a metabolic test, such as the OGTT. In addition, and most importantly, we aimed to develop a new version of TyG for improved performance as marker of OGTT-derived insulin resistance in pregnancy, spanning from normal glucose tolerance to overt GDM. This was accomplished by exploiting advanced methodologies based on machine learning techniques. Notably, to ensure simplicity of the improved TyG, and hence wide applicability in the clinical context, the possible factors assessed as possible complements of TyG for OGTT-derived insulin sensitivity prediction were restricted to those stemming from fasting measures or general subjects’ characteristics.
Discussion
TyG has been shown to be a surrogate marker of insulin resistance [
24,
33,
34]. In this study, we developed an improved version of TyG, for insulin sensitivity assessment in pregnancy. This is particularly important in this population, since impairment in insulin sensitivity is a common trait of pregnancy, even when normal. The new index, called TyGIS, was based on TyG complemented to other parameters. To maintain the simplicity of TyG (likely one of the reasons for its success) we only considered basic patient’s clinical characteristics and variables measured at fasting, thus not requiring metabolic tests such as the OGTT. This choice was also driven by the notion that the OGTT is not always and ubiquitously performed in pregnancy, and when performed the current clinical routine accomplishes a simple three samples, glucose only OGTT, for possible GDM diagnosis. Instead, in our study a more complete OGTT was exploited to compute a reliable parameter of insulin sensitivity (PREDIM), fully validated against the hyperinsulinemic-euglycemic clamp [
59]. Another benefit of our dataset was the heterogeneity of the population, since women had glucose tolerance spanning from normal to GDM, and BMI from lean to overweight or even obese. In addition, the OGTT included glucose, insulin and C-peptide at each time sample, such complete OGTT allowing deep characterization of the TyGIS performance.
We found that TyGIS is a remarkably better surrogate marker than the traditional TyG for insulin sensitivity assessment in pregnancy. Indeed, TyGIS is much strongly associated with the OGTT-derived insulin sensitivity (PREDIM). It also has to be noted that TyGIS was developed as an index of insulin sensitivity, although TyG is an index of insulin resistance. This was done for more direct comparison of TyGIS to its reference, PREDIM, which is an index of insulin sensitivity rather than resistance. We also found that, at difference with TyG, TyGIS agrees with PREDIM as regards determination of insulin sensitivity differences among several subgroups of patients, stratified according to glucose tolerance or BMI. Furthermore, TyGIS calculated at early pregnancy (patients’ visit 1) was good predictor of PREDIM and of abnormal glucose tolerance at later pregnancy stage (second trimester, visit 2), again showing superiority compared to TyG.
It was previously reported that 4.9 mg∙kg
−1∙min
−1, as derived by the clamp (the
M parameter), can be assumed as cut-off for insulin resistance [
73]. PREDIM was developed as predictor of clamp-derived
M parameter, with the same units and expected range of variation [
59]. Similarly, TyGIS was developed as predictor of PREDIM. Thus, the same cut-off value can be reasonably assumed for TyGIS, for determination of insulin resistance. Interestingly, both PREDIM and TyGIS showed several values below that cut-off even in the absence of GDM (details not shown), this being reasonable for women during pregnancy. This result further indicates the reliability of our study findings.
In our study, we also analyzed relationships with model-based parameters of insulin secretion and beta-cell function, this being an important aspect for detailed testing of the new index performance. All indices (TyG, TyGIS and PREDIM) agreed in terms of correlation with insulin secretion parameters (BSR and TIS). All three indices also agreed in showing substantial lack of correlation with the beta-cell function parameters (G-sens, R-sens, PFR). For R-sens and PFR, it has to be noted that in the general population the degree of relationship with insulin sensitivity remains partly unclear, whereas several studies showed that G-sens (the most relevant beta-cell function model parameter) is typically unrelated to insulin sensitivity (this is an advantage, since it implies that there is no need to correct for the insulin sensitivity level for appropriate interpretation [
74‐
79]). Thus, these analyses based on insulin secretion and beta-cell function parameters further proved the reliability of the TyGIS formulation.
In the TyGIS formula, one may expect the presence of BMI. Instead, anthropometry contribution to the new index was provided by simple body weight (BW), as well as lean body mass percentage (LBM, computed by an empirical formula [
67]). Nonetheless, if BW and LBM are replaced by BMI, the performance of the new index deteriorates only slightly (details not shown). The fact that the BMI role resulted to some extent less significant than that of BW and LBM may be due to the reason that in pregnancy the relationship between insulin resistance and BMI may be less strong than in the general population, possibly due to the role of the variable mother’s weight gain, which can act as confounding factor [
18]. Also, insulin resistance in pregnancy depends upon several factors, such as placental hormones and others [
80], and this is another reason for lower relationship between insulin resistance and BMI during gestation. Of note, in the TyGIS formula the BW coefficient is negative, whereas that of LBM is positive. This means that higher BW determines lower TyGIS, as expected (TyGIS is a marker of insulin sensitivity). On the other side, in TyGIS formula, higher LBM determines higher TyGIS, and this again appears reasonable: since LBM is expressed as percentage of total body mass (i.e., BW), for a prescribed BW value higher LBM means lower amount of fat mass, and this is expected to determine better insulin sensitivity.
Opportunities for comparison to previous studies are limited. To our knowledge, only a couple of studies analyzed associations between TyG and insulin resistance in pregnancy [
47,
51]. However, comparison of TyG to insulin resistance indices was limited to that with the fasting HOMA-IR index [
81], thus in pregnancy no studies compared TyG with insulin resistance/sensitivity indices derived by dynamic tests as the OGTT, which typically reflects reference insulin sensitivity (from the clamp) remarkably better than insulin sensitivity at fasting [
82‐
84]. In addition, in pregnancy no studies developed a version of TyG for improved relationship with insulin sensitivity from the OGTT. In Poveda’s study [
47], TyG correlated with HOMA-IR in the first and second trimester of pregnancy, but not in the third trimester, and even in first and second trimesters correlation was weak. This suggests that in pregnancy TyG is not a good surrogate marker of insulin resistance, though on the other hand HOMA-IR has limits in the assessment of insulin resistance, as outlined above [
82‐
84]. Similarly, in the Sánchez-García’s study [
51] the relationship between TyG and HOMA-IR was not strong. Overall, such findings emphasize the need for an improved TyG as surrogate marker of insulin resistance (or sensitivity) in pregnancy. This is in fact what we have done in the present study developing the proposed new index, which on one side is built upon comparison with OGTT-derived insulin sensitivity rather than fasting insulin sensitivity, and on the other side it essentially preserves TyG simple formulation and hence easy applicability in the clinical practice.
Similarly to our study, in previous studies some indices based only on fasting variables were proposed as markers for insulin resistance/sensitivity [
85‐
89]. Specifically, some studies developed an improved version of TyG, but none with pregnant women as the target population [
85‐
87]. In terms of different fasting-based indices, one relatively popular is the triglyceride-HDL ratio (TG/HDL) [
88], or its evolution (SPISE) [
89]. However, after scientific literature analysis we were convinced that TyG performs somehow better than TG/HDL, and this was corroborated in our data by the slightly higher association of TyG rather than TG/HDL to PREDIM, despite such association being weak in both cases (details not shown). Thus, we developed our new index with TyG rather than TG/HDL as the starting point.
TyGIS was derived through machine learning approach. This approach is often used in large datasets, but it is also valuable in smaller datasets with many variables, as in our case [
70,
90‐
93]. Of note, those articles [
70,
90‐
93] emphasize the importance of the initial feature selection process, as we carried out as first step in our approach. It is also worth noting that in the specific domain of pregnancy and GDM we already applied machine learning techniques to a relatively small cohort of former GDM patients to predict the risk of developing type 2 diabetes [
94]. In the machine learning approach of the current study, we first evaluated several variable subsets. Assuming the existence of non-linear dynamics, interactions between variables and their quadratic terms were also considered, but the best model was found to be a linear one. For the selection of the optimal subset, we oriented towards methods providing a simple equation, thus easy to use and hence potentially applicable in the clinical context. Therefore, SVM models were considered, with linear kernel and L2 or L1 regularization, though robust linear regression model was also tested [
95]. Non-linear models were also considered, but they were not providing simple predictor coefficients, and hence did not allow to end up with an algebraic equation of easy interpretation and use [
96]. The next step for final model selection was applying on the training set a nested CV procedure, this yielding unbiased performance estimates even in case of limited dataset size [
70]. The selection of the optimal model was based on both the RMSE obtained from the nested CV and the BIC criteria. The final selected model was that with the lowest BIC and one of the lowest RMSE. Of note, the model with the lowest RMSE had unsatisfactory BIC, and it was in fact not able to discriminate between subgroups on the training set population (details not shown).
Our study has some limitations. The first is the size of the cohort, which was not big. However, this is justified by the wide battery of variables, including OGTT with frequent (30 min) sampling frequency rather than the traditional 1-h frequency, and with measurement of glucose, insulin and C-peptide. Such OGTT characteristics allowed reliable assessment of both insulin sensitivity and insulin secretion/beta-cell function. Indeed, both aspects were relevant for development and testing of the new index performance, since insulin sensitivity and insulin secretion/beta-cell function may be related. Thus, when testing performances of a new insulin sensitivity/resistance index, assessment of its possible relationships with insulin secretion and beta-cell function should definitely be considered. It is also worth noting that datasets larger than the dataset analyzed in the current study, with similar features (variables) available, are rare. This is likely due to the reason that investigators typically limit the burden and discomfort of the study procedures in pregnant women, due to the intrinsic frailty of the gravid condition even in the absence of any morbidity. Thus, to our knowledge few studies (one of which partly from our research group) reported datasets of pregnant women larger than the dataset under analysis with OGTT measurement of glucose, insulin and C-peptide [
97‐
99], and at any rate those studies [
97‐
99] missed lipid profile information (including triglycerides). In summary, we are not aware of any dataset with larger number of participants and all necessary features for the goals of the current study. On the other hand, the size of the current dataset is similar (slightly larger) to that analyzed in some of our previous studies related to pregnancy, where the glucose/insulin/C-peptide OGTT was available [
56,
100‐
102]. Of note, in one of those studies we already applied a technique classified within the machine learning domain (i.e., principal component analysis) [
101].
It also has to be reported that we searched for a precise indication about the appropriate number of samples (observations) in relation to the number of features, but we did not find a unique and robust recommendation. However, for machine learning approaches similar to ours, we found that some investigators suggested a minimum number of observations equal to the number of features plus 50 in cases up to five features, and equal to 10 observations per feature in cases of 6 or more features [
103]. These indications essentially are met in our analysis, since we tested models with two to six features at most (apart for some of the models including the GDM presence variable), and in the training set we had in fact about 10 × 6 observations (precisely, 59). Moreover, if we consider the final selected model, it includes four features only. Thus, the applicable indication for the observations number is that of 50 plus the features number, yielding 54 in our case (this being fully consistent with our training set size). On the other hand, it is worth noting that the main findings of the study confirm the size of the dataset as sufficient for our purposes. Indeed, TyGIS performance was satisfactory according to all tests, in both training and test set. Furthermore, when stratifying the subjects into subgroups, both PREDIM and TyGIS consistently showed the expected differences in insulin sensitivity (i.e., higher in NGT than in non-NGT, and in lean than in overweight people). In addition, as previously outlined, results related to the relationships with model-based insulin secretion and beta-cell function parameters were typically as expected, again according to both PREDIM and TyGIS. It is also worth noting that possible larger datasets would not necessarily improve TyGIS performance further. Indeed, TyGIS by definition does not include information from metabolic tests (such as the OGTT), to maintain the original simplicity and easy use of the traditional TyG. On the other hand, it is unlikely that one index based on variables measured in non-stimulated conditions can predict with extreme accuracy the stimulated metabolic state, even in the case of very large datasets exploited for the index development. For this reason, TyGIS has to be considered a surrogate marker of insulin sensitivity, and this would hold even if TyGIS was developed on larger datasets.
Another limitation of our study was that the reference insulin sensitivity measure, as derived from a hyperinsulinemic-euglycemic clamp, was not available in our dataset. However, the OGTT data allowed calculation of PREDIM, which has same units and range of variation of the clamp-derived index, and was proved being excellent predictor of the latter [
59]. Nonetheless, another issue to consider is whether PREDIM was validated in a sufficiently wide range of insulin sensitivity values, thus comprising values typical of insulin resistance observable in pregnancy, especially when complicated by GDM. To this purpose, it is convenient examining the typical values of insulin sensitivity in pregnancy, as assessed by the reference clamp-derived index. One study reported the clamp
M parameter in both GDM and non-GDM pregnant women, in each trimester of pregnancy [
104]. Lowest average values were observed in GDM in the third trimester (2.70 ± 0.81 mg∙kg
−1∙min
−1, mean ± SD). This translates into 95% interval (mean ± 2 × SD) equal to 1.08–4.32 mg∙kg
−1∙min
−1. However, although 1.08 mg∙kg
−1∙min
−1 is quite low insulin sensitivity value, PREDIM validation included such low values (lowest values ≈0.5 mg∙kg
−1∙min
−1 [
59]). Other studies reported in pregnancy similar clamp
M parameter values [
105,
106]. On the other hand, it should be acknowledged that PREDIM formula includes the BMI, and such variable may be not completely reliable in pregnancy. However, based on our analyses, this potential bias may determine inaccuracies in PREDIM calculation not exceeding 10–15%, which appears acceptable from a clinical point of view (details not shown). It is also worth noting that PREDIM provided reliable findings in some previous studies in pregnancy [
56,
102,
107].
One further limitation was that the lean body mass, which resulted as feature selected for the final TyGIS formula, was not directly measured, but estimated by an empirical formula [
67]. However, the formula is widely accepted and has been extensively used in many investigations. In addition, though not specific for pregnancy, where a contribution to the lean body mass may come from the fetus, the used empirical formula is specific for females, thus we assumed it as sufficiently reliable.
Finally, it is necessary to recommend some restrictions in the use of TyGIS. Indeed, despite TyGIS being remarkably improved compared to the traditional TyG as predictor of insulin sensitivity/resistance, it remains a surrogate marker. Thus, TyGIS is appropriate for possible comparison among different subgroups of pregnant women, but the investigator should be aware that the value in a single subject may be affected by a relevant error. As a fact, whenever possible, insulin sensitivity/resistance should be always assessed stimulating subject’s metabolism with glucose or food administration, as with the OGTT. On the other hand, when the metabolic test is not practically feasible (such as in possible epidemiological studies including large cohorts), TyGIS has proven as good insulin sensitivity surrogate marker, at least in pregnancy.
It may also be asked whether the TyGIS formula could be applied to other populations. In our opinion, even in different populations TyGIS may perform better than TyG, as marker for insulin sensitivity/resistance. However, this was not proved in the present study, thus we do not recommend TyGIS use in other populations. Nonetheless, the current study may trigger future studies aimed at developing improved TyG-based surrogates of insulin sensitivity/resistance specifically tailored for different populations, in the context of precision medicine [
108‐
111]. In such new studies, the methodology of the present study may be exploited as guide for a robust approach. On the other hand, in pregnancy future studies may explore the possible TyGIS ability to predict relevant aspects other than insulin sensitivity, such as maternal/neonatal outcomes.