Predictive serum metabolic markers in GDM patients: A targeted metabolomics study combined with machine learning of sex steroid hormones

Graphical abstract

Highlights

• •Defined serum reference ranges of 20 sex steroid hormones in healthy non-pregnant, healthy pregnant, and GDM women.
• •Elucidated sex steroid hormone profile patterns across the three groups.
• •Revealed sex steroid homeostasis characteristics in GDM subgroups stratified by FBG, OGTT, and BMI.
• •Developed GDM diagnostic models via integration of targeted sex steroid metabolomics and machine learning.

Gestational diabetes mellitus (GDM), characterized by glucose intolerance with onset or first recognition during pregnancy, is a substantial global health burden. It affects approximately 14% of pregnancies and increases the risk of maternal-fetal complications and long-term type 2 diabetes (T2DM) in mothers [1]. Current diagnosis mainly relies on oral glucose tolerance tests (OGTT), which are time-consuming, inconvenient for patients, and have limited predictive value in the early stages of GDM [2]. Accumulating evidence indicates that sex steroid hormones play a crucial role in GDM pathogenesis, as they regulate insulin sensitivity and glucose metabolism, and their dysregulation may contribute to disease development [3]. However, existing studies mostly analyze individual hormones in isolation, neglecting the dynamic changes of hormones and their synergistic effects with clinical indicators, which leads to underutilization of the diagnostic potential of biomarkers [4]. This study employed high-performance liquid chromatography tandem quadrupole mass spectrometry (LC-MS/MS) technology we previously established to quantify characteristic changes in serum sex steroid hormone profiles of GDM patients, screened disease-specific biomarkers via targeted metabolomics, and subsequently established a clinical diagnostic model and validated its diagnostic efficacy using machine learning methods (Fig. 1).

Fig. 1.

Clinical study design and main findings. The study design and key findings of a targeted metabolomics investigation combined with machine learning in patients with gestational diabetes mellitus (GDM) are as follows: integrated analysis reveals disrupted profiles of serum sex steroid hormones in GDM patients, while machine learning identifies estrogen metabolites as promising biomarkers with high diagnostic efficacy. These findings underscore the potential of estrogen metabolites to improve early detection of GDM patients. FBG: fasting blood glucose; BMI: body mass index; NN: neural network; SVM: support vector machine; RF: random forest; 2OHE1: 2-hydroxyestrone; 4OHE2: 4-hydroxyestradiol; 4OHE1: 4-hydroxyestrone; 4MeOE1: 4-methoxyestrone; HbA1C: glycated hemoglobin A1c; ROC: receiver operating characteristic; AUC: area under the curve.

We optimized previous detection methods to quantify 20 serum sex steroid hormones (15 estrogens and 5 androgens) (Fig. S1 and Table S1) [5], with the underlying androgen and estrogen metabolic pathways detailed in Fig. S2. Comprehensive profiling of these 20 hormones was performed on 297 serum samples from three participant groups: 96 healthy non-pregnant women (NC), 101 healthy pregnant women (NP), and 100 GDM patients. Inclusion/exclusion criteria were detailed in Fig. S3, and clinical baseline parameters (age, body mass index (BMI) and basal blood glucose) showed varying degrees of intergroup differences (Table S2).

Fundamental endocrine shifts distinguished the three groups (Figs. S4 and S5, and Tables S3–S5). Androgen concentrations remained stable across pregnancy states, while total estrogens surged more than 60-fold in both NP and GDM groups compared to NC levels, with GDM showing a further elevation relative to NP (Table S3). This estrogen elevation reflected pathway-specific dysregulation: hydroxylated metabolites increased progressively from NC to NP to GDM (particularly the 2-hydroxylated and 4-hydroxylated forms), whereas methylated metabolites showed selective augmentation in GDM compared to NP (Table S4 and S5). Compositionally, androgens dominated NC profiles, while estrogens accounted for over 95% of sex steroids in both pregnant cohorts (Fig. S4). Although NP and GDM groups exhibited nearly identical proportional hormone distributions, analysis of absolute concentrations uncovered a critical divergence: the GDM group had higher total estrogen levels and significantly elevated concentrations of hydroxylated metabolites (Fig. S5). Collectively, these results revealed systematic and significant differences in serum sex steroid hormone profiles across the three groups, with discrepancies between NP and GDM primarily confined to specific estrogen and androgen subclasses.

To clarify the impact of hormone levels on GDM pathophysiological changes, serum hormone data were compared across groups stratified by fasting blood glucose (FBG), OGTT results, and BMI categories (Figs. S6–S8). In FBG subgroups, elevated glucose correlated with reduced total sex hormone concentrations but stable compositional proportions (Fig. S6). Abnormal OGTT shifted estrogen metabolism toward prototype accumulation (with concomitant metabolite reduction) and elevated specific hormones (estrone, estradiol, androstenedione) (Fig. S7). Additionally, advancing maternal obesity progressively intensified estrogen dominance and amplified total hormone concentrations (Fig. S8). Subgroup analyses thus revealed variations in serum sex steroid hormone profiles among GDM patients. When analyzing subgroups based on FBG, OGTT results, or BMI, inter-group comparisons revealed no significant variations in the remaining two key indicators across all subgroup classifications (Tables S6–S8).

Given altered sex steroid hormone homeostasis in pregnant women (further modified in GDM), we used targeted metabolomics to identify GDM-associated key hormones (Fig. S9). Partial least squares discriminant analysis (PLS-DA) and orthogonal projections to latent structures discriminant analysis (OPLS-DA) separated NC from NP and GDM groups but showed less clear distinction between NP and GDM (Figs. S9A–C), with permutation tests confirming no overfitting (Figs. S9D–F). Discriminatory metabolites varied by comparison group: NC vs. NP featured estradiol (E2), estriol (E3), 2-hydroxyestrone (2OHE1), and 2-methoxyestradiol (Fig. S9G); NP vs. GDM identified 2OHE1, 4-hydroxyestradiol, 4-hydroxyestrone, and 4-methoxyestrone (Fig. S9H); NC vs. GDM centered on E2, estrone, E3, and 2-methylestrone (Fig. S9I). Volcano plots validated all differentially expressed metabolites, showing strong concordance with variable importance in the projection (VIP) rankings (Figs. S9J–L). These findings indicated that normal pregnancy involved primary estrogen dysregulation, while GDM was characterized by extensive changes in estrogen metabolic pathways.

We then developed distinct diagnostic models and compared their efficacy to provide novel perspectives for clinical GDM diagnosis. Binary logistic regression constructed four models: Model 1 (7 clinical biochemical parameters); Model 2 (clinical parameters + Biomarker 1, screened from NC vs. GDM); Model 3 (clinical parameters + Biomarker 2, derived from NP vs. GDM); Model 4 (clinical parameters + both biomarkers). The receiver operating characteristic (ROC) curve evaluation showed that gradual inclusion of estrogen metabolites improved diagnostic performance (optimal area under the curve (AUC) 0.944) (Fig. S10A), with model validation (Table S9 and Fig. S10B) and permutation testing (Figs. S10C–F) confirming strong clinical translation potential. Machine learning (neural network (NN), random forest (RF), support vector machine (SVM)) and confusion matrices further validated efficacy (Figs. S11–S13). Statistical analysis showed all three algorithms improved with iterative refinement (Tables S10–S12), with RF identified as the most robust (balanced performance gains, highest final metrics, and most stable parameter tuning).

In conclusion, maternal sex steroid hormones underwent dynamic alterations during pregnancy to meet metabolic demands, but perturbations in estrogen metabolites, particularly in 2OH- and 4OH- pathways, were most pronounced in GDM. Notably, integrating sex steroid hormone levels (especially estrogen metabolites over prototypes) with clinical criteria significantly enhanced GDM diagnostic efficacy. We also elucidated hormone profiles of GDM patients with varying FBG, OGTT, and BMI. The identified estrogen metabolites could serve as valuable adjuncts for GDM clinical screening and diagnosis, potentially improving detection accuracy and timeliness and contributing to better management of this prevalent pregnancy-related condition. This conclusion was based on preliminary evidence from our study and required verification in future research.

CRediT authorship contribution statement

Yi Xu: Formal analysis. Zhulin Gu: Methodology, Conceptualization. Li Wang: Methodology. Xinghan Liu: Software. Hao Wu: Software. Zhenzhou Jiang: Validation, Supervision. Jinfang Song: Software. Changjiang Ying: Validation. Tan Li: Validation. Tao Wang: Project administration, Conceptualization. Na Xie: Formal analysis. Xiaoxing Yin: Funding acquisition, Data curation. Tingting Yang: Conceptualization.

Ethical statement

Ethical approval for the study was obtained from the Ethics Committee of the Affiliated Hospital of Xuzhou Medical University (Approval No.: XYFY2021-KL258). Given the retrospective nature of the research, the requirement for informed consent from individual participants was waived by the Ethics Committee of the Affiliated Hospital of Xuzhou Medical University. These samples were residual specimens from previous clinical diagnosis and treatment, and not collected specifically for this research. Informed consent could not be reasonably obtained as the subjects were unidentifiable and unreachable with excessively high contact costs. All samples were fully anonymized, and the study involved no personal privacy concerns or commercial interests. In addition, all donors had signed informed consent forms with broad authorization for use in all medical research.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the supplementary data to this article: