This retrospective cohort study was conducted at Anhui Maternal and Child Health Hospital, a tertiary maternity hospital in Anhui Province, China. Medical records of pregnant women with HIP who delivered at the hospital between January 2017 and December 2019 were reviewed.

HIP in this study referred to GDM diagnosed during the current pregnancy. Women with known pregestational type 1 or type 2 diabetes were managed in a different clinical pathway and were not included in this cohort. GDM was diagnosed between 24 and 28 weeks of gestation based on a one-step, 75-g oral glucose tolerance test (OGTT) in accordance with the American Diabetes Association Standards of Medical Care in Diabetes, 2022 [1, 3].

For the present analysis, blood biochemistry indices measured at 25–29⁺⁶ weeks of gestation were extracted from the hospital electronic database. A total of 675 women with HIP who delivered singleton live births at $\ge$38 weeks of gestation were included in the study. We focused on term and late-term deliveries to reduce heterogeneity arising from preterm birth-related complications, and to ensure that birthweight reflected fetal growth potential rather than early delivery. Based on neonatal birthweight, participants were classified into an LGA group (n = 113) or a non-LGA group (n = 562). Neonatal weight status was defined according to the INTERGROWTH-21^st international standards for newborn weight by gestational age and sex [7].

The inclusion criteria were: (1) singleton live birth; and (2) gestational age at delivery $\ge$38 weeks. The exclusion criteria were: (1) twin or multiple pregnancy; and (2) incomplete clinical or laboratory information.

Women with HIP were managed at our institution according to national and international guidelines. Initial management consisted of individualized medical nutrition therapy and advice on moderate physical activity. Women were instructed to monitor capillary blood glucose at home. When fasting or postprandial glucose values repeatedly exceeded the target levels (fasting 5.3 mmol/L, 1-hour postprandial 7.8 mmol/L, 2-hour postprandial 6.7 mmol/L), insulin therapy was initiated at the discretion of the treating obstetrician or endocrinologist. However, because complete longitudinal records of self-monitored blood glucose and insulin doses were not consistently available in the electronic database, overall glycemic control during pregnancy could not be formally quantified in this retrospective study.

General characteristics and biochemical indexes were obtained from medical records. The variables included maternal age, height, gravidity, parity, and the following laboratory parameters: total protein (TP), albumin (ALB), globulin (GLO), albumin-to-globulin ratio (A/G), alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin (TBIL), total bile acid (TBA), indirect bilirubin (IBIL), direct bilirubin (DBIL), alkaline phosphatase (ALP), $\gamma$-glutamyl transferase (GGT), lactate dehydrogenase (LDH), low-density lipoprotein cholesterol (LDL), high-density lipoprotein cholesterol (HDL), triglycerides (TG), total cholesterol (CHOL), uric acid (UA), creatinine (CREA), blood urea nitrogen (BUN), cystatin C (CysC), creatine kinase (CK), $\alpha$-hydroxybutyrate dehydrogenase (HBDH), homocysteine (HCY), and fasting plasma glucose (GLU). In total, 29 candidate predictors were considered.

All biochemical measurements were performed on fasting venous blood samples in the same hospital laboratory following standard operating procedures. Continuous variables were summarized as the median (P25–P75) and compared between groups using the Wilcoxon rank-sum test. Categorical variables were expressed as n (%) and compared using the chi-square test or Fisher’s exact test, as appropriate. A two-sided p-value 0.05 was considered statistically significant.

Information on pre-pregnancy weight, body mass index (BMI), gestational weight gain, and glycated hemoglobin (HbA1c) was not systematically recorded in the electronic database and therefore could not be included as candidate predictors in this analysis.

Pearson correlation analyses were performed to explore collinearity among candidate predictors. When two variables were highly correlated (absolute correlation coefficient $>$0.7), the variable with stronger clinical interpretability and/or a stronger univariable association with LGA was retained and the other was excluded to reduce multicollinearity. The linearity of each continuous predictor with the outcome logit was then examined using restricted cubic splines.

Subsequently, univariable logistic regression was used to examine the association between each candidate predictor and LGA. Predictors with p 0.10 in the univariable analyses, together with clinically important factors, were entered into a multivariable logistic regression model. Stepwise selection was applied to obtain the final model and reported odds ratio (OR) with 95% confidence interval (CI). To evaluate the models’ robustness, multicollinearity was determined using variance inflation factors (VIFs), and influential observations were identified with Cook’s distance. Using regression coefficients from the final model, we constructed a nomogram to estimate the probability of delivering an LGA infant among women with HIP. Cases with missing values in any candidate predictor were excluded. With 113 LGA events and 8 predictors retained, the events-per-variable ratio exceeded 10, indicating adequate statistical power for logistic regression.

Model discrimination was evaluated by calculating the area under the receiver operating characteristic curve (AUC). Calibration was assessed using the Brier score, the Hosmer-Lemeshow goodness-of-fit test, and calibration plots that compared predicted and observed risks across deciles of predicted probability. For internal validation, we ran bootstrap resampling with 1000 repetitions and reported optimism-corrected performance estimates. Clinical usefulness was examined by decision curve analysis (DCA), which quantifies the net benefit of applying the model across a range of threshold probabilities relative to treating all patients or none. All analyses were conducted in R (version 4.1.3; R Foundation for Statistical Computing, Vienna, Austria). Model development and validation were reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) statement.

The correlation coefficients between A/G and GLO, TP and GLO, AST and ALT, IBIL and TBIL, HBDH and LDH, and CHOL and LDL were all $>$0.7 (Fig. 1). With each highly correlated pair, the variable that was more clinically interpretable or more strongly associated with LGA was retained, while the other variable was excluded (A/G, TP, AST, TBIL, HBDH, and LDL). This was carried out to reduce multicollinearity.

Fig. 1.

Pairwise correlation matrix of candidate predictors included in the model development. Darker colors indicate stronger positive or negative correlations. ALB, albumin; GLO, globulin; A_G, albumin-to-globulin ratio; TP, total protein; ALT, alanine aminotransferase; AST, aspartate aminotransferase; TBIL, total bilirubin; TBA, total bile acid; IBIL, indirect bilirubin; DBIL, direct bilirubin; ALP, alkaline phosphatase; GGT, $\gamma$-glutamyl transferase; LDH, lactate dehydrogenase; LDL, low-density lipoprotein cholesterol; HDL, high-density lipoprotein cholesterol; TG, triglycerides; CHOL, total cholesterol; UA, uric acid; CREA, creatinine; BUN, blood urea nitrogen; CysC, cystatin C; CK, creatine kinase; HBDH, $\alpha$-hydroxybutyrate dehydrogenase; HCY, homocysteine; GLU, fasting plasma glucose.

The results of the univariable analysis comparing the two groups are shown in Table 1. Among the general characteristics, only maternal height was significantly different between the LGA and non-LGA groups (p 0.05). For the biochemical indices, TG and GLU levels were significantly higher in the LGA group compared with the non-LGA group, whereas ALB, GLO, HDL, CHOL, BUN, and HCY levels were all significantly lower (p 0.05). No significant differences between the two groups were observed for the remaining laboratory parameters.

The Pearson correlation test was conducted on 29 independent variables. The ‘rms’ package was employed on the R 4.1.3 platform to fit a regression model via stepwise regression. The final mathematical equation derived from the multivariable logistic regression model to predict LGA infants was: logit(p) = –9.5305 – 0.1564 $\times$ ALB (g/L) + 0.0100 $\times$ LDH (IU/L) + 0.2201 $\times$ TG (mmol/L) – 0.2551 $\times$ CHOL (mmol/L) – 0.4320 $\times$ BUN (mmol/L) – 0.1922 $\times$ HCY (µmol/L) + 0.3008 $\times$ GLU (mmol/L) + 0.0856 $\times$ height (cm) (Table 2). To present the prediction model, we constructed a nomogram that allows quantitative estimation of the risk of LGA in infants born to women with HIP (Fig. 2). Additionally, we developed a website with a calculation tool (https://China-ahsfybjy-model.shinyapps.io/dynnomapp/). Restricted cubic splines were used to explore potential non-linear relationships between each continuous predictor and the log odds of LGA (Fig. 3). For all predictors in the final model, the p-values for the non-linear components were $>$0.05, indicating no strong statistical evidence of departure from linearity within the observed ranges. The variables were also evaluated using the VIF and Cook’s distance. The VIF of the models was significantly below 10 (range 1.0395–1.0976), which suggests there was no covariance among the variables. The point-of-influence analysis demonstrated that Cook’s distance of the models was considerably 1, indicating minimal influence of abnormal values in these variables (Fig. 4).

Fig. 2.

Nomogram for predicting the risk of delivering a large-for-gestational-age infant in women with hyperglycemia in pregnancy based on maternal height and mid-pregnancy biochemical indicators.

Fig. 3.

Restricted cubic spline plots showing the relationship between each continuous predictor and the log odds of large-for-gestational-age birth. No strong evidence of non-linearity was observed.

Fig. 4.

Cook’s distance plots for the final logistic regression model, indicating the influence of individual observations.

We assessed discrimination using the area under the receiver operating characteristic (ROC) curve. The nomogram achieved an AUC of 0.7333 (95% CI: 0.6856–0.7810), indicating moderate discriminatory ability (Fig. 5). Model calibration was primarily evaluated using the calibration plot, which showed close agreement between the predicted and observed probabilities of LGA across deciles of predicted risk (Fig. 6). The Brier score was 0.1260, indicating acceptable overall accuracy. The Hosmer Lemeshow goodness-of-fit test yielded $\chi$² = 14.7410 (df = 8, p = 0.0644), providing no strong evidence of lack of fit. However, this test has important sample size-related limitations and should thus only be interpreted as supplementary information to the graphical assessment.

Fig. 5.

Receiver operating characteristic (ROC) curve of the prediction model in the original dataset. The shaded area represents the 95% confidence interval for the AUC. AUC, area under the receiver operating characteristic curve.

Fig. 6.

Calibration plot comparing predicted and observed risks of large-for-gestational-age birth across deciles of predicted probability in the original dataset.

DCA was used to assess the model’s clinical effectiveness, with the results shown in Fig. 7. Application of the nomogram prediction model yields a higher net benefit for women with HIP when the predicted probability ranges from 0.0 to 0.65.

Fig. 7.

Decision curve analysis of the prediction model, showing net benefit across a range of threshold probabilities compared with treat-all and treat-none strategies.

Bootstrap resampling was performed 1000 times to validate the model. The ROC curve after resampling (Fig. 8) shows an AUC of 0.7132 (95% CI: 0.7052–0.7329). The Brier score after resampling was calculated as 0.1302. The calibration chart after resampling is depicted in Fig. 9. Hence, the model demonstrates acceptable predictive capability.

Fig. 8.

ROC curve of the model after 1000 bootstrap resamples for internal validation.

Fig. 9.

Calibration plot of the model after 1000 bootstrap resamples for internal validation.

Multiple logistic regression (MLR) is widely used in health outcomes research, with applications in medicine, statistics, and machine learning (computer science). MLR involves fitting multiple parameters in a predictive model by assuming a linear or additive relationship between predictors and outcomes [11]. In the current study, we built a predictive model for LGA using general information and biochemical indicators for HIP patients extracted from hospital electronic medical records. The AUC of this model was 0.7333 (95% CI: 0.6856–0.7810), indicating moderate discriminative ability. To validate the model’s accuracy, we performed bootstrap validation 1000 times. The AUC value of our model after bootstrapping was 0.7132 (95% CI: 0.7052–0.7329), demonstrating acceptable predictive performance and reasonable stability. Univariate analysis results revealed statistically significant differences in height, ALB, GLO, HDL, TG, CHOL, BUN, HCY, and GLU between the LGA and non-LGA groups. Furthermore, multivariate logistic regression analysis indicated that height, ALB, LDH, TG, CHOL, BUN, HCY, and GLU are significant predictors of LGA risk in HIP women.

Some authors have proposed that LGA may be linked to factors such as pre-pregnancy overweight and obesity, excessive weight gain during pregnancy, abnormal glucose metabolism, abnormal lipid metabolism, and genetic factors [12, 13, 14]. The results of multivariate regression analysis indicated that maternal height (OR = 1.09, p 0.001) was a risk factor for LGA in HIP patients. Children born to taller HIP women had higher birth weights, which aligns with findings from numerous studies conducted in various countries [15, 16, 17]. Miletić et al. [18] suggested that the uterine environment of shorter mothers is smaller, thereby limiting the space available for intrauterine growth. It is well known that the weight of pregnant women can vary, whereas their height remains constant. Therefore, maternal height may be a better predictor of fetal birth weight than maternal weight [19].

Several studies have emphasized the importance of fasting blood glucose (FBG) as a screening tool for GDM and as a predictor of adverse neonatal outcomes [20]. A meta-analysis revealed a positive association between elevated fasting glucose levels in OGTT trials and adverse pregnancy outcomes, including LGA [21]. Fetal glucose exposure and subsequent insulin secretion can lead to excessive conversion of glucose to fat, resulting in fetal weight gain [22]. This finding is consistent with results from the present study.

Substantial literature exists on the relationship between blood lipid levels and GDM. However, there is still some uncertainty regarding the reported results on abnormal blood lipid status in diabetic pregnant women [23]. Some studies have found that women with GDM have elevated triglycerides, LDL, and total cholesterol (CHOL), while HDL levels are decreased [24, 25]. A meta-analysis of 4168 women with GDM and 9718 healthy pregnant women showed that compared to women with standard glucose tolerance, those with GDM had higher triglyceride levels and lower HDL levels. However, CHOL and LDL levels were similar between the two groups [26]. Our study also found that high triglyceride levels are a risk factor for LGA in women with HIP (OR = 1.25, p = 0.005), in line with the findings of Shi et al. [27, 28, 29]. A possible reason could be that excess triglycerides in the mother are absorbed, metabolized by the placenta, and transported to the fetus in various forms, leading to excessive fetal growth. However, it remains unclear whether abnormal lipid levels in the blood directly affect fetal body mass.

Additionally, our study found that CHOL was a protective factor against LGA in women with HIP (OR = 0.77, p = 0.011). In contrast, Kaneko et al. [30] found that among Japanese mothers without GDM who delivered full-term babies, lower and higher maternal CHOL levels in the second trimester were associated with small-for-gestational-age (SGA) and LGA infants, respectively. In contrast, Shi et al. [27] found no significant correlations between CHOL or LDL concentrations and perinatal outcomes, including LGA, in either GDM or non-GDM women. These varying research findings may be attributed to pre-pregnancy differences in BMI, age, diet, region, and race. Our study also identified LDH as a risk factor for LGA in women with HIP (OR = 1.01, p = 0.010). There is still only limited domestic and international research on the relationship between LDH and LGA. Increased LDH activity indicates metabolic disturbances in glucose utilization and lactate production, which can lead to diabetic complications [31]. This may explain why elevated LDH levels contribute to LGA in pregnant women with HIP.

Human serum albumin (HSA) plays a crucial role in various physiological functions and is vital for maintaining good health. Hypoalbuminemia occurs due to a combination of inflammation and insufficient nutritional intake, and is commonly observed in patients with chronic diseases. Inflammation is closely associated with vascular diseases and can result in damage to the vascular endothelium [32]. Clinical and biomedical research has indicated a positive regulatory cycle between hyperglycemia and interstitial inflammation. Hyperglycemia triggers the synthesis of proinflammatory cytokines, chemokines, and adipokines in both the placenta and peripheral blood, ultimately affecting fetal growth [33]. In the present study, higher serum albumin levels were inversely associated with the risk of LGA in women with HIP (OR = 0.86, p = 0.002). To our knowledge, this specific association has not been reported previously, although study has suggested that albumin may modulate systemic inflammation and oxidative stress [34]. It is therefore plausible that better-preserved albumin levels reflect a more favorable metabolic and inflammatory profile, which might contribute to a lower risk of LGA. However, this finding should be interpreted with caution and confirmed in future studies.

Several authors have investigated the association between GDM and homocysteine (HCY), but the findings have so far been inconsistent. Liu et al. [35] found that HCY levels in women with GDM were significantly higher than those in women with standard glucose tolerance. On the other hand, Radzicka et al. [36] found no significant correlation between HCY levels and glucose tolerance. In a study of 7587 participants from the Ottawa and Kingston birth cohorts, maternal HCY concentration was found to be associated with a higher incidence of placental-mediated complications, including SGA, preeclampsia, and placental abruption [37]. Our research found similar results, suggesting that HCY acts as a protective factor against LGA (OR = 0.83, p = 0.004). This may be attributed to the involvement of HCY in inflammatory reactions, leading to dysfunction of maternal and placental vascular endothelial cells [38], which in turn affects fetal growth and development.

Urea nitrogen is a byproduct of protein metabolism and is excreted by the kidneys. Previous studies have shown that urea nitrogen is an independent risk factor for GDM [39]. However, it remains unclear whether BUN (blood urea nitrogen) can cause LGA. The current study found that BUN was a protective factor for LGA in pregnant women with HIP (OR = 0.65, p = 0.009). A possible explanation is that the hyperglycemic state leads to glycosylation of plasma and tissue proteins, impairing their normal function and reducing the excretion of urea nitrogen by the kidneys. Consequently, the level of BUN increases, and excessive amounts enter the placenta, thereby affecting the growth and development of the fetus.

In our model, FBG measured at 25–29⁺⁶ weeks was retained as an independent predictor of LGA. Although fasting glucose is part of the diagnostic workup for HIP, the measurement used in this study reflects glycemic status at mid-pregnancy, following the diagnosis and start of lifestyle management or insulin therapy in many women, rather than the diagnostic OGTT value per se. We were unable to include HbA1c as an additional marker of longer-term glycemic control because it was not routinely measured in all pregnant women with HIP during the study period. Future studies with more complete information on HbA1c and longitudinal glucose profiles may help to clarify the relative contributions of chronic versus short-term glycemic control to the risk of LGA.

HCY emerged as a significant predictor in our final model. HCY is measured as part of routine biochemical testing during mid-pregnancy at our hospital, thus facilitating its inclusion in the present analysis. However, testing for HCY is not universally available in all antenatal care settings. This may limit the direct transferability of our nomogram to some institutions, and also highlights the need to evaluate simplified models that rely only on universally available predictors.

Our study has several limitations. First, because of the challenges in obtaining comprehensive biochemical data during pregnancy, the sample size was limited to 675 women, and the number of LGA events was modest. We therefore used bootstrap resampling rather than a split-sample approach for internal validation, and external validation in independent cohorts is still required. Second, the analysis was restricted to women who delivered at $\ge$38 weeks of gestation. This design choice reduced heterogeneity related to preterm birth, but may have systematically excluded some higher-risk HIP pregnancies with earlier delivery. This potentially limits the ability to generalize our model to preterm or early-term births.

Third, important clinical factors such as pre-pregnancy BMI, gestational weight gain, HbA1c, and detailed longitudinal information on self-monitored blood glucose and insulin use were not consistently available in the electronic database and could not be incorporated. Consequently, we were unable to formally quantify overall glycemic control during pregnancy or evaluate its interaction with the biochemical predictors, which may have led to residual confounding. Fourth, although several predictors in the final model (albumin, BUN, HCY) are biologically linked to endothelial dysfunction and hypertensive disorders of pregnancy, we could not reliably ascertain the occurrence of preeclampsia or other hypertensive disorders in this retrospective cohort. The observed associations should therefore be interpreted with caution, as they may partially reflect underlying placental pathology.

Fifth, we focused exclusively on LGA and did not analyze SGA or fetal growth restriction (FGR) as competing fetal growth outcomes. Some of the predictors associated with a lower risk of LGA in our model may in fact be markers of placental dysfunction and increased risk of SGA/FGR. Future studies with larger sample sizes are warranted to simultaneously evaluate predictors of both excessive and restricted fetal growth in women with HIP. Finally, to handle multicollinearity, we used a simple, correlation-based strategy to choose between pairs of highly correlated predictors. Although this approach is transparent and easy to implement, it may not be optimal from a predictive modeling perspective. Instead, penalized regression methods such as LASSO or elastic net could be explored in future work to further refine variable selection.