Results

Characteristics of study population

In this study, a comprehensive analytical framework was established to evaluate the predictive value of first-trimester laboratory markers for PE, with the overall study design and workflow summarized in Figure 1. Of the 12,715 pregnancies in the study population, PE occurred in 556 cases (4.4%). The characteristics of the study population are summarized in Figure 2 and Table S1. Compared with the non-PE group, the PE group exhibited higher maternal age and BMI and a higher prevalence of twin gestation and IVF conception, along with a lower parity. Before the 14th week of gestation, the PE group showed significantly higher levels of liver function-related indices, including TBA, PALB, GLB, ALT, ALP, GGT, LDH, and CHE, and lower levels of A/G, TBIL, DBIL, and IBIL. Additionally, renal function-related indices such as the UAR MoM, along with lipid- and inflammation-related indices, such as TyG MoM, TC, LDL-C, VLDL-C, and CRP, were significantly elevated in the PE group. In contrast, HDL-C levels were decreased in the PE group. Furthermore, micronutrient-related indices such as Na, K, Ca, and Cu were elevated, while CO2, Mg, and Fe levels were reduced in the PE group.

Figure 1: Flowchart of the study. This study included 12,715 pregnancies (with 140 cases of abnormal pregnancy excluded), among which 556 cases developed PE. Clinical data (age, BMI, plurality, parity, IVF, etc.), ultrasound parameters (UtA-PI), and laboratory markers (liver function, blood lipids, inflammation, trace elements, etc.) were collected before 14 weeks of gestation. Multivariate logistic regression, RCS analysis, and two-stage linear regression models were used to evaluate the association and threshold effects between laboratory markers and PE. Subsequently, the preterm PE and term PE cohorts were randomly divided into training and testing sets at an 8:2 ratio, respectively. The Boruta algorithm and seven machine learning models were employed for feature selection and the construction of laboratory markers prediction models. The optimal model was selected based on AUC and sensitivity, and SHAP values were used to quantify the contribution of features. Next, the prediction score of the optimal laboratory markers prediction model was integrated with the FMF model (maternal risk factors, MAP, PLGF, and UtA-PI) for the construction of the predictive modeling of PE, and the prediction performance was evaluated. Finally, cfDNA sequencing and TSS profiling were performed on 51 maternal plasma samples (18 cases of preterm PE and 33 controls), revealing early metabolic dysfunction and heightened inflammation in cases destined for PE. PE, preeclampsia; BMI, body mass index; IVF, in vitro fertilization; UtA-PI, uterine artery pulsatility index; RCS, restricted cubic spline; AUC, area under the receiver operating characteristic curve; SHAP, shapley additive explanations; FMF, fetal medicine foundation; MAP, mean arterial pressure; PLGF, placental growth factor; cfDNA, cell-free DNA; TSS, transcription start site; GGT, gamma-glutamyltransferase; ALT, alanine aminotransferase; ALP, alkaline phosphatase; CHE, cholinesterase; FBG, fasting blood glucose; UAR MoM, uric acid-albumin ratio multiple of the median; TC, total cholesterol; TyG MoM, triglyceride-glucose index multiple of the median; CRP, c-reactive protein; Cu, copper; Ca, calcium; Mg, magnesium; Fe, iron; CatBoost, categorical boosting; LR, logistic regression.

Figure 2: Laboratory markers changes in early pregnancy in PE pregnant women. Differences in (A) TBA, (B) PALB, (C) ALT, (D) ALP, (E) GGT, (F) LDH, (G) CHE, (H) UAR MoM, (I) TyG MoM, (J) TC, (K) HDL-C, (L) LDL-C, (M) VLDL-C, (N) CRP, and (O) Mg levels between PE cases and controls. PE, preeclampsia; TBA, total bile acids; PALB, prealbumin; ALT, alanine aminotransferase; ALP, alkaline phosphatase; GGT, gamma-glutamyltransferase; LDH, lactate dehydrogenase; CHE, cholinesterase; UAR MoM, uric acid-albumin ratio multiple of the median; TyG MoM, triglyceride-glucose index multiple of the median; TC, total cholesterol; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; VLDL-C, very low-density lipoprotein cholesterol; CRP, C-reactive protein; Mg, magnesium. ∗∗∗p < 0.001.

Association analysis of routine laboratory markers and risk of PE

In order to exclude the interference of confounding factors and reveal the true independent associations between various laboratory markers and PE risk, the study used three models to gradually adjust for potential confounding variables (Fig. 3; Table S2). In Model 3, which included adjustments for all confounding variables, liver function-related indices, including PALB (OR = 1.01, 95% CI, 1.01–1.02, p < 0.001), GLB (OR = 1.03, 95% CI, 1.01–1.05, p = 0.046), ALT (OR = 1.01, 95% CI, 1.01–1.01, p = 0.008), ALP (OR = 1.02, 95% CI, 1.02–1.03, p < 0.001), GGT (OR = 1.03, 95% CI, 1.02–1.03, p < 0.001), LDH (OR = 1.01, 95% CI, 1.01–1.01, p = 0.002), and CHE (OR = 1.01, 95% CI, 1.01–1.01, p < 0.001), were found to be significantly positively associated with PE prevalence, and DBIL (OR = 0.89, 95% CI, 0.80–0.99, p = 0.043) was significantly negatively associated with PE risk. Similarly, renal function-related indices, such as the UAR MoM (OR = 10.16, 95% CI, 7.33–14.09, p < 0.001), as well as lipid- and inflammation-related indices, including TyG MoM (OR = 4.29, 95% CI, 3.40–5.42, p < 0.001), TC (OR = 1.17, 95% CI, 1.05–1.30, p = 0.005), LDL-C (OR = 1.39, 95% CI, 1.22–1.57, p < 0.001), VLDL-C (OR = 1.86, 95% CI, 1.25–2.77, p = 0.002), and CRP (OR = 1.05, 95% CI, 1.03–1.06, p < 0.001), were also positively associated with PE prevalence. In contrast, HDL-C levels (OR = 0.48, 95% CI, 0.37–0.63, p < 0.001) were significantly inversely associated with PE risk. Additionally, micronutrient-related indices, such as Ca (OR = 4.03, 95% CI, 1.86–8.72, p < 0.001) and Cu (OR = 1.03, 95% CI, 1.01–1.05, p = 0.001), were positively associated with PE prevalence, whereas CO2 (OR = 0.96, 95% CI, 0.93–0.99, p = 0.021), Mg (OR = 0.07, 95% CI, 0.02–0.27, p < 0.001), and Fe (OR = 0.97, 95% CI, 0.96–0.99, p < 0.001) levels were inversely associated with PE risk. However, TBA, A/G, TBIL, IBIL, Na, and K were not found to be significantly associated with PE risk.

Figure 3: Association between liver function-related indices in early pregnancy and the risk of PE. Model 1: unadjusted model; Model 2: adjusted for age, BMI, and plurality; Model 3: adjusted for age, BMI, plurality, parity, and IVF conception. PE, preeclampsia; BMI, body mass index; IVF, in vitro fertilization; OR, odds ratio; CI, confidence interval; PALB, prealbumin; GLB, globulin; ALT, alanine aminotransferase; ALP, alkaline phosphatase; GGT, gamma-glutamyltransferase; LDH, lactate dehydrogenase; CHE, cholinesterase; DBIL, direct bilirubin.

Quartile analysis reveals dose-effect trends

To investigate the potential nonlinear relationship between laboratory routine indicators and PE, as well as to identify high-risk populations, continuous laboratory markers were categorized into quartiles. As shown in Figure 3, Figure 4A and Table S2, when compared with the first quartile, a significantly higher prevalence of PE was observed in the fourth quartile of liver function-related indices, including TBA, PALB, GLB, ALT, ALP, GGT, LDH, and CHE, with increases of 0.54, 1.82, 0.4, 0.96, 0.83, 1.92, 0.51, and 2.27, respectively. However, the prevalence of PE was lower in the fourth quartile of IBIL levels.

Figure 4: The proportion of patients with PE sorted by quartiles. (A) The proportion of patients with PE sorted by quartiles of liver function-related indices and renal function-related indicators (TBA, PALB, GLB, ALT, ALP, GGT, LDH, CHE, IBIL, and UAR MoM). (B) The proportion of patients with PE sorted by quartiles of lipid-, inflammation- and micronutrient-related indicators (TyG MoM, TC, HDL-C, LDL-C, VLDL-C, CRP, Ca, Mg, Cu, and Fe). PE, preeclampsia; TBA, total bile acids; PALB, prealbumin; GLB, globulin; ALT, alanine aminotransferase; ALP, alkaline phosphatase; GGT, gamma-glutamyltransferase; LDH, lactate dehydrogenase; CHE, cholinesterase; IBIL, indirect bilirubin; UAR MoM, uric acid-albumin ratio multiple of the median; TyG MoM, triglyceride-glucose index multiple of the median; TC, total cholesterol; HDL-C, high-density lipoprotein cholesterol; LDL-C, low-density lipoprotein cholesterol; VLDL-C, very low-density lipoprotein cholesterol; CRP, c-reactive protein; Ca, calcium; Mg, magnesium; Cu, copper; Fe, iron.

Similarly, PE prevalence was significantly higher in the fourth quartile of renal function-related indices, such as UAR MoM (Fig. 4A), as well as in lipid- and inflammation-related indices, including TyG MoM, TC, LDL-C, VLDL-C, and CRP (Fig. 4B). Conversely, a significantly lower incidence of PE was found in the highest quartile of HDL-C (Fig. 4B). Additionally, an increased incidence of PE was noted in the fourth quartile of micronutrient-related indices, such as Ca and Cu, whereas a reduction in PE incidence was observed in the fourth quartile of Mg and Fe (Fig. 4B). This trend remained significant after adjusting for potential confounders (Table S2).

Exploring the nonlinear relationship and threshold effects of laboratory markers in preeclampsia risk

To further ensure the reliability of the findings and identify potential threshold effects between laboratory routine indicators and PE risk, a nonlinear relationship analysis was conducted using an RCS regression model. After adjusting for all confounders, nonlinear associations with PE risk were observed for TBA, ALT, ALP, GGT, CHE, TyG MoM, HDL-C, CRP, Mg, Cu, and Fe (P for nonlinearity < 0.05) (Fig. 5). In contrast, linear relationships with PE risk were identified for A/G, PALB, GLB, LDH, TBIL, DBIL, IBIL, UAR MoM, TC, LDL-C, VLDL-C, Na, K, Ca, and CO2 (P for nonlinearity > 0.05).

Figure 5: The nonlinear relationship between laboratory markers and PE risk. Restricted cubic spline analyses the association of PE with laboratory markers: (A) TBA; (B) ALT; (C) ALP; (D) GGT; (E) CHE; (F) TyG MoM; (G) HDL-C; (H) CRP; (I) Mg; (J) Cu; (K) Fe. PE, preeclampsia; TBA, total bile acids; ALT, alanine aminotransferase; ALP, alkaline phosphatase; GGT, gamma-glutamyltransferase; CHE, cholinesterase; TyG MoM, triglyceride-glucose index multiple of the median; HDL-C, high-density lipoprotein cholesterol; CRP, c-reactive protein; Mg, magnesium; Cu, copper; Fe, iron.

Subsequently, the threshold effects of laboratory routine indicators were examined to assess potential non-linear associations with the risk of PE. As shown in Table S3, threshold effects were identified for TBA, ALT, GGT, and CRP, with respective thresholds of 2.59 μmol/L, 37 U/L, 35 U/L, and 8.09 mg/L. Below these thresholds, each one-unit increase in the corresponding marker was associated with an elevated risk of PE by 0.24-fold (OR = 1.24; 95% CI, 1.05–1.45; p = 0.009), 0.03-fold (OR = 1.03; 95% CI, 1.02–1.04; p < 0.001), 0.06-fold (OR = 1.06; 95% CI, 1.04–1.07; p < 0.001), and 0.18-fold (OR = 1.18; 95% CI, 1.13–1.24; p < 0.001), respectively. However, no significant associations were observed above these thresholds. Similarly, threshold values were determined for ALP (41 U/L), CHE (254 U/L), TyG MoM (0.83), and Cu (31.39 μmol/L). Below these cut-off points, no significant associations with PE risk were observed. In contrast, above these cut-off values, each one-unit increase in the respective indicators was associated with increased PE risk by 0.03-fold (OR = 1.03; 95% CI, 1.02–1.03; p < 0.001), 0.01-fold (OR = 1.01; 95% CI, 1.01–1.01; p < 0.001), 5.06-fold (OR = 6.06; 95% CI, 4.50–8.17; p < 0.001), and 0.09-fold (OR = 1.09; 95% CI, 1.04–1.14; p < 0.001), respectively. For HDL-C, Mg, and Fe, the threshold values were 1.84 mmol/L, 0.91 mmol/L, and 25.1 μmol/L, respectively. Below these thresholds, each one-unit increase was associated with a reduced risk of PE by 0.70-fold (OR = 0.30; 95% CI, 0.18–0.51; p < 0.001), 0.98-fold (OR = 0.02; 95% CI, 0.00–0.13; p < 0.001), and 0.05-fold (OR = 0.95; 95% CI, 0.93–0.97; p < 0.001), respectively. No significant associations were identified above these cut-off thresholds.

Laboratory markers predict pre-symptomatic PE

Based on these findings, the study further utilized machine learning to evaluate the predictive efficacy of these laboratory indicators. As previous studies have indicated that FMF models vary in predictive efficacy for different subtypes of PE, such as preterm and term PE, we sought to determine whether routine laboratory markers measured in early pregnancy could predict these PE subtypes.

For preterm PE, 22 out of 26 differentially expressed features were identified as significant by the Boruta algorithm and were selected for model development. These 22 features were input into seven machine learning models, and the optimal combination of hyperparameters for preterm PE prediction was determined through 10-fold cross-validation. Model performance was assessed using sensitivity, AUC, and F1 score on the testing set. The AdaBoost model demonstrated both the highest AUC (0.741) along with superior sensitivity (66.8%), and thus was considered the most effective among all models tested (Fig. 6A and B). To facilitate interpretability of the AdaBoost model, SHAP analyses were conducted at both global and individual levels. The SHAP summary plot identified UAR MoM, TyG MoM, PALB, GGT, and CRP as key contributors to AdaBoost model predictions. High values of these variables were associated with increased PE risk (Fig. 6C). In an individual case, elevated CRP, PALB, and TyG MoM increased predicted risk, whereas low UAR MoM reduced it, consistent with global feature importance (Fig. 6D).

Figure 6: Laboratory markers predict pre-symptomatic PE. (A) The ROC curves of the preterm PE prediction models constructed by seven machine learning algorithms in the training set. (B) The ROC curves of the preterm PE prediction models constructed by seven machine learning algorithms in the testing set. (C) The SHAP summary plot of the AdaBoost model shows the effect of laboratory markers on the preterm PE prediction model. (D) Waterfall plot demonstrates the impact of specific laboratory markers on the predicted risk of preterm PE. (E) The ROC curves of the term PE prediction models constructed by seven machine learning algorithms in the training set. (F) The ROC curves of the term PE prediction models constructed by seven machine learning algorithms in the testing set. (G) The SHAP summary plot of the LightGBM model shows the effect of laboratory markers on the term PE prediction model. (H) Waterfall plot demonstrates the impact of specific laboratory markers on the predicted risk of term PE. PE, preeclampsia; ROC, receiver operating characteristic; SHAP, shapley additive explanations.

For term PE, the Boruta algorithm identified 19 of the 26 differential features as important, which were subsequently used for model development. Among all models evaluated, LightGBM demonstrated the best overall performance in the testing set for screening purposes. It achieved a sensitivity of 53.8%, an AUC of 0.689, and a balanced accuracy of 68.2%, with minimal evidence of overfitting (Fig. 6E and F). To enhance the interpretability of the LightGBM model, SHAP analyses were performed at both population and individual levels. The SHAP summary plot revealed TyG MoM, PALB, GGT, and CRP as the most influential predictors in the LightGBM model. Elevated levels of TyG MoM and PALB were associated with increased PE risk, while higher GGT and Ca levels demonstrated protective effects (Fig. 6G). In a representative case analysis, increased TyG MoM and PALB values contributed positively to the risk prediction, whereas reduced GGT and elevated Ca levels decreased the predicted risk, aligning with the overall feature importance trends (Fig. 6H).

Integration of routine laboratory markers and maternal risk profiles for preeclampsia early prediction

Among the cohort, 5901 pregnant women underwent PE risk prediction during pregnancy using the FMF model, which incorporated maternal risk factors as well as MAP, PLGF, and UtA-PI. This prompted an investigation into whether routine laboratory markers could be integrated with the aforementioned parameters to enhance prediction performance through machine learning approaches.

For preterm PE, the highest AUC (0.881) and sensitivity (60.9%) were achieved by the Catboost model in the testing set, making this the most effective among the models evaluated (Fig. 7A–C).

Figure 7: Integration of laboratory markers and maternal risk profiles to predict PE. (A) The ROC curves of the preterm PE prediction models in the training set. (B) The ROC curves of the preterm PE prediction models in the testing set. (C) Confusion matrix analysis of CatBoost model for preterm PE. For the confusion matrix: X-axis (actual class): Label “0″ denotes women clinically diagnosed without PE (non-PE cases) and label “1″ denotes women clinically diagnosed with PE (PE cases); Y-axis (predicted class): Label “0″ indicates a model prediction of non-PE (low risk cases) and label “1″ indicates a model prediction of PE (high risk cases). (D) The ROC curves of CatBoost and FMF model for preterm PE in the entire cohort. (E) The ROC curves of the term PE prediction models in the training set. (F) The ROC curves of the term PE prediction models in the testing set. (G) Confusion matrix analysis of logistic regression model for term PE. (H) The ROC curves of logistic regression and FMF model for term PE in the entire cohort. PE, preeclampsia; ROC, receiver operating characteristic; FMF, fetal medicine foundation.

When applied to the entire study population, Catboost's AUC rose to 0.954 (95% CI, 0.914–0.969), a marked improvement over the established FMF model (AUC = 0.854; 95% CI, 0.799–0.857; p < 0.001) (Fig. 7D). Moreover, at a prespecified false-positive rate of 14%, the Catboost model detected 92.5% of preterm PE cases, compared with just 65.0% sensitivity for the FMF algorithm. These findings underscore the potential of machine learning-derived risk stratification to substantially enhance early identification of women at highest risk for preterm PE, particularly in settings where the incremental gain in sensitivity could translate into timely prophylactic interventions and improved maternal–fetal outcomes.

For term PE, the logistic regression model demonstrated the highest AUC (0.891) and sensitivity (86.7%) in the testing set, and thus was considered the most effective among all models assessed (Fig. 7E–G). When applied to the entire cohort, the AUC of the logistic regression model increased to 0.913 (95% CI, 0.887–0.931), which greatly exceeded the AUC of the FMF algorithm of 0.778 (95% CI, 0.738–0.810; p < 0.001) (Fig. 7H). At a fixed false-positive rate of 14%, the logistic model identified 81.1% of term PE cases, whereas the sensitivity of the FMF model was only 47.7%. These results suggest that a simplified regression-based approach is promising to augment existing FMF screening and thereby significantly improve the early prediction of term PE.

Early transcriptomic alterations reveal multiple-system dysregulation in preeclampsia

To elucidate the mechanisms underlying early laboratory markers changes in pregnancy, low-coverage whole-genome sequencing was applied to compare cfDNA coverage across TSSs between cases and controls. Reduced promoter coverage inversely reflected higher gene expression in maternal or placental tissues.9,10,15 This allowed for a noninvasive differential expression analysis, which revealed early transcriptional alterations associated with pregnancies that later developed PE.

In total, 51 maternal plasma samples underwent low-depth cfDNA sequencing and TSS profiling, including 18 cases diagnosed with preterm PE—a subtype strongly associated with adverse pregnancy outcomes. Compared with controls, women who later developed preterm PE exhibited significant early pregnancy alterations in GGT and the TyG MoM (Fig. 8A and B). Three datasets were downloaded from GEO, and 803 DEGs were identified in placentas from women with PE using the thresholds of P.adj < 0.05 and |log2 fold change| > 0.2 (Fig. 8C and D). In mid-pregnancy cfDNA analysis, 627 DEGs were obtained under the P.adj < 0.05 criterion. Intersection analysis revealed 26 overlapping genes, suggesting that TSS signals can reflect early functional changes in the placenta during pregnancy. In particular, expression of the FLNB gene was markedly up-regulated in placental tissue, consistent with prior reports,16 whereas it was down-regulated at the TSS level, representing a concordant pattern. Because approximately 10% of cfDNA is derived from the placenta, with the majority originating from maternal leukocytes and the liver, cfDNA analysis may also capture functional alterations in these maternal tissues and organs. Functional enrichment using GSEA revealed coordinated disruptions in key biological processes, including hepatic metabolic dysfunction, lipid processing abnormalities, and systemic inflammatory activation, all of which aligned with routine laboratory markers perturbations observed in early pregnancy in individuals who subsequently developed PE.

Figure 8: Early transcriptomic alterations reveal multiple-system dysregulation in pregnancies that develop PE. (A) Compared with control, women who later developed preterm PE had significantly higher GGT in early pregnancy. (B) Compared with control, women who later developed preterm PE had significantly higher TyG MoM in early pregnancy. (C) Volcano plot of DEGs (|log2 FC| > 0.2, P.adj < 0.05). The red indicates up-regulated genes and the blue indicates down-regulated genes between PE and control groups in the GEO datasets. (D) Intersection of GEO-derived placental tissue DEGs and mid-pregnancy cfDNA-derived DEGs. (E) GSEA analysis showed significant negative enrichment of pathways related to hepatic function (metabolism of xenobiotics by cytochrome P450, aspirin metabolism, drug metabolism cytochrome P450, glucuronidation, phase II conjugation of compounds) in PE cases. (F) GSEA analysis showed significant negative enrichment of pathways related to bile acid synthesis and clearance (disorders of bile acid synthesis and biliary transport, synthesis of bile acids and bile salts via 7 alpha hydroxycholesterol) in PE cases. (G) GSEA analysis showed significant negative enrichment of pathways related to lipid oxidation (oxysterols derived from cholesterol, fatty acid omega oxidation) in PE cases. (H) GSEA analysis showed significant up-regulation of pro-inflammatory signaling (IL-6 pathway, IL-1 structural pathway) in PE cases. (I) GSEA analysis showed that starch and sucrose metabolism pathways were significantly down-regulated, and insulin signaling pathways were significantly up-regulated in PE cases. (J) Differential promoter coverage of functionally relevant genes between control and preterm PE cases. PE, preeclampsia; GGT, gamma-glutamyltransferase; TyG MoM, triglyceride-glucose index multiple of the median; DEGs, differentially expressed genes; GSEA, gene set enrichment analysis; cfDNA, cell-free DNA; TSS, transcription start site. ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001.

Gene sets related to cytochrome P450-mediated metabolism, glucuronidation, aspirin metabolism, drug metabolism via cytochrome P450, and phase II conjugation enzymes were significantly negatively enriched (Fig. 8E). This broad suppression of hepatic enzymatic pathways was interpreted as indicative of reduced maternal capacity to metabolize endogenous waste products and oxidative metabolites. As a result, the accumulation of toxic lipid intermediates, uremic solutes, and pro-inflammatory mediators may occur, imposing additional stress on the maternal vasculature and placenta. The increased cfDNA representation of ADH7 and GLYATL3, reflecting reduced gene expression in contributing tissues, was consistent with clinical observations of elevated liver enzymes (e.g., ALT, GGT, ALP) during early pregnancy in individuals who later developed PE (Fig. 8J).

Impaired detoxification also appeared to affect bile acid synthesis and clearance. Notably, the biliary transport pathway and the synthesis of bile acids and bile salts via the 7α-hydroxycholesterol pathway were negatively enriched, implicating genes such as AKR1C4 and AKR1D1 as playing critical roles in bile acid biosynthesis and cholesterol elimination (Fig. 8F, J). The repression of these pathways may contribute to the dyslipidemic phenotype frequently observed in PE.

Similarly, down-regulation of the oxysterols derived from the cholesterol pathway (NES = −2.22) and the fatty acid ω-oxidation pathway (NES = −1.57) was interpreted as indicative of impaired lipid oxidation and cholesterol derivative metabolism (Fig. 8G). Genes such as ADH7 and AKR1C4 were consistently up-regulated in cfDNA profiles, reflecting reduced expression in contributing tissues and suggesting dysfunctional energy metabolism in both trophoblasts and the maternal liver (Fig. 8J). These metabolic disruptions may lead to lipid accumulation, placental lipotoxicity, and the systemic dyslipidemia characteristic of PE.

In contrast, pro-inflammatory signaling was markedly up-regulated in PE cases. Significant enrichment of the interleukin-6 (IL-6) signaling pathway (NES = +1.95) and the interleukin-1 (IL-1) structural signaling pathway (NES = +2.00) was observed (Fig. 8H). These pathways featured key immune regulators, including CEBPB, FOXO1, and MAP3K3 (Fig. 8J). The activation of these pathways suggested a systemic inflammatory state during early pregnancy, likely promoting endothelial activation, immune cell infiltration, and placental dysfunction.

Metabolic pathway analysis revealed concurrent evidence of glucose handling defects and insulin signaling dysregulation. The starch and sucrose metabolism pathway (NES = −2.51; p < 0.001) was significantly down-regulated in PE cases, with key enzymes such as SI involved in glycolysis and carbohydrate interconversion found to be transcriptionally repressed (Fig. 8I and J). These findings suggest reduced glucose utilization and impaired carbohydrate flux, consistent with early insulin resistance. Conversely, the insulin signaling pathway was significantly up-regulated (NES = +2.06; p < 0.001), with core enriched genes including IGF1R, IRS2, and FOXO3 mediating canonical insulin signaling (Fig. 8I and J). Although this transcriptional up-regulation may initially suggest enhanced pathway activity, it more likely reflects a compensatory response to impaired insulin sensitivity.