Maternal circulating lipid metabolic heterogeneity in preeclampsia: machine learning identifies clinically distinct subtypes Authors: Jianfeng Gan, Li Cao, Jiejie Shao, Weirong Gu, Rong Hu, Yu Xiong, Yinan Wang, Huanqiang Zhao, Qiongjie Zhou, Xiaotian Li Published: 2026 ABSTRACT BackgroundPreeclampsia is associated with dyslipidemia. Maternal circulating lipid profiles may reflect underlying placental-metabolic interactions. We aimed to identify distinct maternal circulating lipid metabolic subtypes and assess their association with clinical outcomes among preeclamptic women.MethodsA retrospective cohort study of 1,124 women with preeclampsia (2018–2022) was conducted. Third-trimester serum lipids (total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C) and low-density lipoprotein cholesterol (LDL-C)) were analyzed. Unsupervised k-means clustering (Euclidean distance, k = 4 optimized by consensus culmulative distribution function) was used to identify subtypes. Clinical characteristics and pregnancy outcomes were compared across clusters.ResultsWe identified four lipid metabolic subtypes of preeclampsia with distinct clinical profiles. Four subtypes were identified: Cluster 1 (low metabolic subtype, LMS, 33.5%), Cluster 2 (high HDL subype, HHS, 35.0%), Cluster 3 (high triglyceride subtype, HTS, 11.7%), and Cluster 4 (high metabolic subtype, HMS, 19.8%). The HTS subtype with elevated triglycerides and highest pre-pregnancy body mass index (median 23.92 kg/m2), had the lowest rates of preterm birth (3.1%, p = 0.016) and low birth weight (3.8%, p < 0.001). Conversely, the HMS subtype exhibited the highest blood pressure and proteinuria levels.ConclusionsThis study establishes lipid-driven subtypes of preeclampsia with distinct clinical trajectories, suggesting placental lipid handling may be a key modulator in disorder heterogeneity. This heterogeneity underscores the importance of lipid-based stratification for understanding placental dysfunction in preeclampsia.Graphical Abstract Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12884-026-09527-x. Introduction Preeclampsia, a pregnancy-specific disorder originating from placental dysfunction, leads to systemic metabolic disturbances including dyslipidemia [1]. Affecting 2–8% of pregnancies globally, it significantly increases maternal and fetal morbidity and mortality [2–4], with long-term cardiovascular implications for both mother and offspring [5–7]. The clinical heterogeneity of preeclampsia, likely driven by diverse placental pathologies [8, 9], complicates the development of targeted interventions [10]. The placenta plays a central role in lipid metabolism during pregnancy, regulating cholesterol transport and triglyceride oxidation [11]. Dyslipidemia, which ischaracterized by elevated total cholesterol (TC), triglycerides (TG), low-density lipoprotein (LDL), and reduced high-density lipoprotein (HDL), is a hallmark of preeclampsia [12, 13]. Notably, HDL dysfunction may exacerbate placental oxidative stress while paradoxically exhibiting protective effects in some subtypes [14–16]. Despite these associations, maternal circulating contribution to lipid metabolic heterogeneity remains poorly understood. Recent studies highlight distinct lipid profiles in early- versus late-onset preeclampsia [17, 18], suggesting divergent pathogenic mechanisms. For example, early-onset cases often exhibit severe dyslipidemia linked to impaired placental vascular remodeling [19], whereas late-onset cases may reflect maternal metabolic syndrome [20]. However, traditional classifications based on clinical phenotypes fail to capture this complexity. Unsupervised clustering offers a data-driven approach to identify subtypes by integrating multi-dimensional lipid data [21]. Prior studies have not applied this method to classify maternal circulating lipid metabolic subtypes or correlate them with clinical outcomes. Our study addresses this gap and aims to: (1) identify lipid-based subtypes in preeclampsia using k-means clustering; (2) evaluate their association with placental dysfunction markers; and (3) propose possible links between subtype-specific lipid profiles and placental pathophysiology. Methods Study design and participants This retrospective cohort study analyzed pregnant women who received prenatal care and delivered at the Obstetrics and Gynecology Hospital of Fudan University (January 2018–August 2022). Inclusion criteria were: (1) based on the diagnosis of preeclampsia per ACOG guidelines [22]: new-onset hypertension (systolic blood pressure (BP) ≥ 140 mmHg or diastolic BP ≥ 90 mmHg) after 20 weeks gestation with either proteinuria (≥ 300 mg/24-hour urine collection) or end-organ dysfunction (e.g., thrombocytopenia, renal insufficiency), and (2) availability of third-trimester lipid profiles. Exclusion criteria were multiple pregnancies, chronic or gestational hypertension without preeclampsia, and missing delivery or lipid data. The procedure of enrollment is illustrated in Fig. 1. Fig. 1DAG indicating the relationship between lipid metabolism, preeclampsia and potential confounding factors.HDP: hypertensive disorders during pregnancy; BMI: body mass index. DAG indicating the relationship between lipid metabolism, preeclampsia and potential confounding factors.HDP: hypertensive disorders during pregnancy; BMI: body mass index. Ethical approval was obtained from the Ethics Committee of the Obstetrics and Gynecology Hospital of Fudan University (2021-144), and the requirement for informed consent was waived due to the de-identification of data. The study adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. Data collection Participants were identified through the hospital’s electronic medical records (EMR) system. Trained researchers extracted de-identified data, including: (1) demographic information: maternal age, education level (categorized as bachelor’s degree or above/high school or below), smoking/alcohol use; (2) clinical measurement: pre-pregnancy body mass index (BMI), gestational weight gain, blood pressure, proteinuria, gestational diabetes mellitus (GDM); and (3) pregnancy outcomes: delivery mode (vaginal delivery and Cesarean delivery), placental weight, neonatal birth weight, and complications (e.g., neonatal asphyxia, defined as Apgar score < 7 at 5 min). Maternal BP was measured twice at admission using calibrated sphygmomanometers (mean values recorded). Proteinuria was quantified via 24-hour urine collection (mg/24hr). Postpartum hemorrhage was defined as blood loss ≥ 500 mL (vaginal delivery) or ≥ 1,000 mL (Cesarean delivery). Placental weight was measured immediately after delivery trimmed of cord and membranes. Maternal serum lipid measurement The exposure was maternal serum lipid profile in the third trimester, including total cholesterol (TC), triglycerides (TG), high-density lipoprotein cholesterol (HDL-C) and low-density lipoprotein cholesterol (LDL-C). Sample collection was from fasting venous blood drawn between 28 and 42 weeks at the regular prenatal visit, centrifuged at 3,000 rpm for 10 min, and analyzed within 2 h. An automatic biochemical analyzer (LABOSPECT 008α, Japan) was utilized to detect blood lipid levels, including TC (LabAssay Cholesterol, Wako Pure Chemical Co., Ltd., Osaka, Japan), TG (LabAssay Triglyceride, Wako), HDL-C (HDL-Cholesterol, Sekisui), LDL-C (Cholestest LDL, Sekisui Medical Technology Co., Ltd., Tokyo, Japan). The standard reference ranges used for third-trimester pregnancy in our institution are: TC < 6.5 mmol/L, TG < 3.0 mmol/L, LDL-C < 3.5 mmol/L, HDL-C > 1.0 mmol/L. Quality control was based on duplicate measurements by two technicians with coefficients of variation < 5%. Outcomes measurement Outcome variables included preeclampsia with severe features per ACOG guideline [22, 23], preterm birth, low birth weight and very low birth weight. Detailed definition and diagnosis criteria were showed in Table S1. Specifically, preeclampsia with severe features is defined as a subset of patients demonstrating severe hypertension and/or specific signs or symptoms indicative of significant end-organ dysfunction; preterm birth is defined as delivery occurring before 37 completed weeks of gestation; low birth weight (LBW) is defined as birth weight less than 2500 g and very low birth weight (VLBW) is defined as birth weight less than 1500 g. Statistical analyses Data description and preprocessing Continuous variables were presented as mean ± standard deviation (SD) or median (interquartile range) for non-normally distributed data, while categorical variables were expressed as numbers (percentages). Maternal serum lipid concentrations were summarized using 2.5th-97.5th percentiles to describe their distributions. Prior to clustering analysis, extreme outliers (values beyond the 1st and 99th percentiles for each lipid parameter) were winsorized to the respective percentile boundaries to reduce the influence of extreme values. After winsorization, all lipid variables (TC, TG, LDL, HDL) were standardized using Z-score transformation (mean = 0, SD = 1) to eliminate scale differences between measurements. Unsupervised clustering Euclidean distance was selected for the k-means algorithm due to its sensitivity to magnitude differences in lipid profiles, for identifying subtypes based on absolute lipid levels. K-means clustering via the ConsensusClusterPlus package (R v4.0.3) (reps = 50, pItem = 0.8, pFeature = 1, clusterAlg="km", distance="euclidean”, seed = 888), were performed. Optimal cluster number was determined by evaluating solutions for k = 2 to 9. The following quantitative metrics were used: (1) plateau in the cumulative distribution function (CDF) curve: the curve showed an elbow and reached a plateau at k = 4 (Fig. 3); (2) maximized average silhouette width (intra-cluster cohesion): the average silhouette width for k = 4 was 0.43, the highest among k = 2-9 (Figure S1); (3) Cluster stability: bootstrap resampling (100 iterations) was performed, yielding a mean Jaccard similarity index indicating high reproducibility (Figure S2); and (4) clinical interpretability of subtypes (LMS, HHS, HTS, HMS). Cluster validation was confirmed by Z-score distributions for each lipid parameter were visualized relative to cluster centers. Pairwise Spearman’s rank correlation coefficients were calculated to assess inter-lipid relationships within clusters. Statistical analysis Continuous variables was compared by one-way ANOVA with Bonferroni post-hoc correction for multiple comparisons (normally distributed data) or Kruskal-Wallis test (non-parametric data), and categorical variables by Pearson’s χ² test or Fisher’s exact test (expected counts < 5). No imputation was performed due to the unsupervised nature of the analysis. Logistic regression to evaluate associations between lipid subtypes (reference: Cluster 3 [HTS]) and adverse outcomes. Multivariable modeling was applied: (1) Model 1: adjusted for maternal age (minimal adjustment); (2) Model 2: adjusted for maternal age + pre-pregnancy BMI + GDM. Since the relationships of metabolic syndrome-related indicators, including pre-pregnancy. To account for potential confounders, three multivariable models were constructed based on a directed acyclic graph (DAG; Fig. 1). BMI, gestational weight gain, and GDM, with lipid metabolism and adverse pregnancy outcomes are established. According to causal inference principles, these variables can act as both confounders and colliders (mediators) on the pathway between lipid subtypes and outcomes. Therefore, we established two adjustment models with prespecified baseline covariates considered clinically important or statistically significant. Due to educational level data having a missing rate exceeding 70%, this variable was excluded from the final analysis. Analyses were performed using R software (v4.0.3; R Foundation) for clustering and correlation analyses, and SPSS (v26.0; IBM Corp.) for regression models. A two-tailed p-value < 0.05 was considered statistically significant. Results Participant characteristics A total of 25,967 pregnant women giving birth during the study period. Among them, 1,681 (6.47%) individuals were diagnosed with preeclampsia. After excluding 297 cases of multiple pregnancies and 240 cases with incomplete or missing information of maternal serum lipid, a total of 1,124 preeclamptic women were included for the unsupervised clustering analysis (Fig. 2). The cohort had a mean maternal age of 32.79 ± 4.58 years and pre-pregnancy BMI of 23.12 kg/m2 (IQR: 20.71–25.92). Comorbidities included GDM (16.0%) and chronic hypertension (0.4%). To assess the potential impact of excluding cases with missing lipid data, we compared baseline characteristics between included (n = 1,124) and excluded (n = 240) women. No significant differences were observed in maternal age, BMI, GDM, or blood pressure (Table S1). Fig. 2Participant enrollment flowchart Participant enrollment flowchart Identification of lipid metabolic subtypes Unsupervised clustering of Z-score-standardized lipid profiles (TC, TG, LDL, HDL) revealed four distinct subtypes (Fig. 3): (1) low metabolic subtype (LMS): characterized by uniformly low TC, TG, LDL, HDL, n = 377 (33.5%); (2) high HDL subtype (HHS): characterized by elevated HDL (+ 1.2 SD) with reduced TG (− 0.8 SD), n = 393 (35.0%); (3) high TG subtype (HTS): characterized by high TG (+ 1.5 SD) and TC (+ 1.1 SD) but relatively low LDL/HDL, n = 131 (11.7%); and (4) high metabolic subtype (HMS): characterized by globally elevated lipid levels, n = 223 (19.8%). The subtypes was data-driven phenotypes, and the average silhouette width was 0.43 indicating moderate cluster cohesion. TC positively correlated with TG, LDL and HDL (all p < 0.01) except in HHS (TC-HDL: p = 0.21). HDL showed a strong negative correlation with TG in HHS (r = − 0.62, p < 0.001). Fig. 3Determination of optimal cluster number using k-means clustering. A Consensus clustering matrix demonstrating the stability of four-cluster partitioning. Darker squares indicate higher co-clustering probabilities. B Cumulative distribution function (CDF) curve for cluster solutions (k=2-9). The plateau at k=4 (arrow) indicates optimal cluster stability. Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS, Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS). Determination of optimal cluster number using k-means clustering. A Consensus clustering matrix demonstrating the stability of four-cluster partitioning. Darker squares indicate higher co-clustering probabilities. B Cumulative distribution function (CDF) curve for cluster solutions (k=2-9). The plateau at k=4 (arrow) indicates optimal cluster stability. Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS, Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS). Clinical profiles and pregnancy outcomes by subgroups The demographic and clinical phenotypes of the four subgroups exhibited significant heterogeneity (Table 1). Notably, pre-pregnancy BMI and GDM differed significantly across clusters (p < 0.001 and p = 0.007, respectively). The highest BMI was recorded in Cluster 3 (median: 23.92, IQR: 21.51–26.92), while the lowest was in Cluster 4 (median: 22.52, IQR: 19.91–24.63). The prevalence of GDM was also highest in LMS (20.7%) and lowest in HMS (11.3%). No significant differences were noted in maternal age, smoking and alcohol consumption, gestational weight gain, previous history of hypertension among the four subgroups. Their clinical characteristics were differed significantly, including maternal systolic blood pressure (p = 0.010), diastolic BP (p = 0.001) and proteinuria (p < 0.001). Notably, maternal BP and proteinuria levels were highest in HMS and lowest in HTS. The four lipid subtypes demonstrated distinct clinical profiles: LMS was characterized by globally low lipids; HHS by high HDL and low TG; HTS by high TG and TC but lower LDL and/or HDL, alongside the highest BMI and GDM rate but the lowest BP and proteinuria; and HMS by hyperlipidemia, the highest BP and proteinuria, and an intermediate risk profile. Table 1Baseline and clinical characteristics of participants across the clustersAll participantsn = 1124Cluster 1n = 377Cluster 2n = 393Cluster 3n = 131Cluster 4n = 223P valueMaternal age, years32.79 ± 4.5833.05 ± 4.6832.53 ± 4.6433.18 ± 4.7532.57 ± 4.160.274Maternal alcohol consumption4 (0.4)1 (0.3)1 (0.3)1 (0.8)1 (0.4)0.726Maternal smoking4 (0.4)1 (0.3)1 (0.3)1 (0.8)1 (0.4)0.726Pre-pregnancy BMI, kg/m223.12(20.71, 25.92)23.41(21.02, 26.42)23.01(20.72, 25.64)23.92(21.51, 26.92)22.52(19.91, 24.63)< 0.001Gestational weight gain, kg14.22 ± 6.5014.10 ± 6.2314.22 ± 6.6114.72 ± 6.6113.91 ± 6.680.789Previous history of hypertension4 (0.4)0 (0)3 (0.8)1 (0.8)0 (0)0.123Previous history of adverse pregnancy107 (9.5)31 (8.2)42 (11)11 (8.4)23 (10)0.634Systolic blood pressure, mmHg132.64 ± 13.80131.85 ± 13.60132.09 ± 13.63131.79 ± 12.02135.45 ± 15.070.010Diastolic blood pressure, mmHg86.06 ± 10.3485.96 ± 9.8285.06 ± 10.5685.19 ± 8.5488.49 ± 11.430.001Proteinuria, g/24 hours1.41 ± 2.111.19 ± 1.681.41 ± 1.921.10 ± 1.802.01 ± 3.03< 0.001GDM180 (16.0)78 (20.7)52 (13.2)24 (18.3)26 (11.3)0.007Data were presented as mean (SD), n (%), or median (IQR)BMI body mass index, GDM gestational diabetes mellitus Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) Table 1Baseline and clinical characteristics of participants across the clustersAll participantsn = 1124Cluster 1n = 377Cluster 2n = 393Cluster 3n = 131Cluster 4n = 223P valueMaternal age, years32.79 ± 4.5833.05 ± 4.6832.53 ± 4.6433.18 ± 4.7532.57 ± 4.160.274Maternal alcohol consumption4 (0.4)1 (0.3)1 (0.3)1 (0.8)1 (0.4)0.726Maternal smoking4 (0.4)1 (0.3)1 (0.3)1 (0.8)1 (0.4)0.726Pre-pregnancy BMI, kg/m223.12(20.71, 25.92)23.41(21.02, 26.42)23.01(20.72, 25.64)23.92(21.51, 26.92)22.52(19.91, 24.63)< 0.001Gestational weight gain, kg14.22 ± 6.5014.10 ± 6.2314.22 ± 6.6114.72 ± 6.6113.91 ± 6.680.789Previous history of hypertension4 (0.4)0 (0)3 (0.8)1 (0.8)0 (0)0.123Previous history of adverse pregnancy107 (9.5)31 (8.2)42 (11)11 (8.4)23 (10)0.634Systolic blood pressure, mmHg132.64 ± 13.80131.85 ± 13.60132.09 ± 13.63131.79 ± 12.02135.45 ± 15.070.010Diastolic blood pressure, mmHg86.06 ± 10.3485.96 ± 9.8285.06 ± 10.5685.19 ± 8.5488.49 ± 11.430.001Proteinuria, g/24 hours1.41 ± 2.111.19 ± 1.681.41 ± 1.921.10 ± 1.802.01 ± 3.03< 0.001GDM180 (16.0)78 (20.7)52 (13.2)24 (18.3)26 (11.3)0.007Data were presented as mean (SD), n (%), or median (IQR)BMI body mass index, GDM gestational diabetes mellitus Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) Baseline and clinical characteristics of participants across the clusters 23.12 (20.71, 25.92) 23.41 (21.02, 26.42) 23.01 (20.72, 25.64) 23.92 (21.51, 26.92) 22.52 (19.91, 24.63) Data were presented as mean (SD), n (%), or median (IQR) BMI body mass index, GDM gestational diabetes mellitus Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) Pregnancy outcomes by subgroups Across the four clusters, their pregnancy delivery week, delivery mode, placental weight and neonatal sex were similar (p > 0.05). The incidence of preterm birth (p = 0.016) and low birth weight (p < 0.001) were of significant difference, while preeclampsia with severe features (p = 0.216) and very low birth weight were of no difference (p = 0.079). Additionally, other adverse pregnancy outcomes including the incidence of postpartum hemorrhage, placental abruption, neonatal asphyxia, neonatal respiratory distress syndrome, was of no significant difference (p > 0.05). The incidence of preterm birth and low birth weight varied significantly across clusters. The incidence of preterm delivery was highest in HHS at 13.0% and lowest in HTS at 3.1%; the proportions of LBW, and VLBW were highest in HHS (19.8% and 5.9%, respectively) and lowest in HTS (3.8% and 0.8%, respectively) (Table 2). Table 2Comparison of pregnancy outcomes across the clustersAll participantsn = 1124Cluster 1n = 377Cluster 2n = 393Cluster 3n = 131Cluster 4n = 223P valueGestational age at delivery, weeks38.09 ± 2.3638.04 ± 2.4037.95 ± 2.4738.70 ± 1.3138.08 ± 2.520.105Delivery mode0.584 Cesarean delivery601 (53.5)209 (55.4)204 (51.9)74 (56.5)114 (51.1) Vaginal delivery523 (47.5)168 (44.6)189 (48.1)57 (43.5)109 (48.9)Postpartum hemorrhage56 (5.0)24 (6.4)19 (4.8)4 (3.1)9 (4.0)0.392Placental abruption14(1.2)5 (1.3)5 (1.3)2 (1.5)2 (0.9)0.979Birth weight, g3058.65 ± 686.403128.9 ± 704.512927.6 ± 701.043329.16 ± 545.593011.66 ± 645.55< 0.001Placental weight, g505.01 ± 162.71509.31 ± 121.80496.65 ± 220.21512.62 ± 85.85509.88 ± 117.530.809Neonatal sex0.595 Male561 (49.9)180 (47.7)191 (48.6)68 (51.9)122 (54.7) Female562 (50.0)197 (52.3)202 (51.4)63 (48.1)101 (45.3)Neonatal asphyxia169 (15.0)64 (17.0)60 (15.3)16 (12.2)29 (13.0)0.446NRDS44 (3.9)14 (3.7)18 (4.6)2 (1.5)10 (4.5)0.443Preeclampsia with severe features354 (31.5)126 (33.4)125 (31.8)31 (23.3)72 (32.3)0.216Preterm birth117 (10.4)39 (10.3)51 (13.0)4 (3.1)23 (10.3)0.016Low birth weight168 (14.9)50 (13.3)78 (19.8)5 (3.8)35 (15.7)< 0.001Very low birth weight47 (4.2)15 (4.0)23 (5.9)1 (0.8)8 (3.6)0.079NRDS: Neonatal respiratory distress syndromeData were presented as mean (SD), n (%), or median (IQR)Preterm birth is defined as delivery occurring before 37 completed weeks of gestationLow birth weight (LBW) is defined as birth weight less than 2500 gVery low birth weight (VLBW) is defined as birth weight less than 1500 gCluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) Table 2Comparison of pregnancy outcomes across the clustersAll participantsn = 1124Cluster 1n = 377Cluster 2n = 393Cluster 3n = 131Cluster 4n = 223P valueGestational age at delivery, weeks38.09 ± 2.3638.04 ± 2.4037.95 ± 2.4738.70 ± 1.3138.08 ± 2.520.105Delivery mode0.584 Cesarean delivery601 (53.5)209 (55.4)204 (51.9)74 (56.5)114 (51.1) Vaginal delivery523 (47.5)168 (44.6)189 (48.1)57 (43.5)109 (48.9)Postpartum hemorrhage56 (5.0)24 (6.4)19 (4.8)4 (3.1)9 (4.0)0.392Placental abruption14(1.2)5 (1.3)5 (1.3)2 (1.5)2 (0.9)0.979Birth weight, g3058.65 ± 686.403128.9 ± 704.512927.6 ± 701.043329.16 ± 545.593011.66 ± 645.55< 0.001Placental weight, g505.01 ± 162.71509.31 ± 121.80496.65 ± 220.21512.62 ± 85.85509.88 ± 117.530.809Neonatal sex0.595 Male561 (49.9)180 (47.7)191 (48.6)68 (51.9)122 (54.7) Female562 (50.0)197 (52.3)202 (51.4)63 (48.1)101 (45.3)Neonatal asphyxia169 (15.0)64 (17.0)60 (15.3)16 (12.2)29 (13.0)0.446NRDS44 (3.9)14 (3.7)18 (4.6)2 (1.5)10 (4.5)0.443Preeclampsia with severe features354 (31.5)126 (33.4)125 (31.8)31 (23.3)72 (32.3)0.216Preterm birth117 (10.4)39 (10.3)51 (13.0)4 (3.1)23 (10.3)0.016Low birth weight168 (14.9)50 (13.3)78 (19.8)5 (3.8)35 (15.7)< 0.001Very low birth weight47 (4.2)15 (4.0)23 (5.9)1 (0.8)8 (3.6)0.079NRDS: Neonatal respiratory distress syndromeData were presented as mean (SD), n (%), or median (IQR)Preterm birth is defined as delivery occurring before 37 completed weeks of gestationLow birth weight (LBW) is defined as birth weight less than 2500 gVery low birth weight (VLBW) is defined as birth weight less than 1500 gCluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) Comparison of pregnancy outcomes across the clusters NRDS: Neonatal respiratory distress syndrome Data were presented as mean (SD), n (%), or median (IQR) Preterm birth is defined as delivery occurring before 37 completed weeks of gestation Low birth weight (LBW) is defined as birth weight less than 2500 g Very low birth weight (VLBW) is defined as birth weight less than 1500 g Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) When comparing the four subgroups, HTS remained associated with the lowest risk of adverse neonatal outcomes. In the logistic regression analysis, the crude odds ratios (OR) for preterm birth, LBW, and VLBW were significantly elevated in LMS, HHS and HMS (Table 3; Fig. 4). After adjusting for maternal age and GDM, these associations remained significant. Notably, HHS exhibited the most pronounced changes, with OR for preterm birth, LBW, and VLBW being 4.788 (95% CI 1.693–13.535, p = 0.003), 6.271 (95% CI 2.478–15.870, p < 0.001), and 7.991 (95% CI 1.068–59.806, p = 0.043), respectively. In addition, risk of preeclampsia with severe features was not statistically significant by subgroups. Table 3Association between different clusters and adverse pregnancy outcomesN (%)P valueCrude modelAdjusted model1Adjusted model2OR(95%CI)P valueOR(95%CI)P valueOR(95%CI)P valuePreeclampsia with severe features354 (31.5)0.216Cluster 1126 (33.4)1.619(1.026–2.555)0.0381.636(1.034–2.586)0.0351.645(1.04–2.601)0.033Cluster 2125 (31.8)1.505(0.954–2.372)0.0791.553(0.983–2.455)0.0591.532(0.969–2.423)0.068Cluster 331 (23.3)Ref-Ref-Ref-Cluster 472 (32.3)1.538(0.941–2.513)0.0861.588(0.969–2.601)0.0661.560(0.952–2.558)0.078Preterm birth117 (10.4)0.016Cluster 139 (10.3)3.663(1.283–10.459)0.0153.687(1.291–10.533)0.0153.701(1.295–10.573)0.015Cluster 251 (13.0)4.735(1.677–13.367)0.0034.853(1.717–13.714)0.0034.788(1.693–13.535)0.003Cluster 34 (3.1)Ref-Ref-Ref-Cluster 423 (10.3)3.651(1.234–10.804)0.0193.742(1.263–11.084)0.0173.680(1.242–10.906)0.019Low birth weight168 (14.9)< 0.001Cluster 150 (13.3)3.853(1.502–9.884)0.0053.871(1.508–9.932)0.0053.891(1.516–9.986)0.005Cluster 278 (19.8)6.240(2.468–15.776)< 0.0016.352(2.511–16.072)< 0.0016.271(2.478–15.870)< 0.001Cluster 35 (3.8)Ref-Ref-Ref-Cluster 435 (15.7)4.691(1.789–12.3)0.0024.773(1.819–12.523)0.0014.693(1.788–12.318)0.002Very low birth weight47 (4.2)0.079Cluster 115 (4.0)5.387(0.705–41.185)0.1055.386(0.704–41.177)0.1055.416(0.708–41.414)0.104Cluster 223 (5.9)8.081(1.081–60.436)0.0428.072(1.079–60.396)0.0427.991(1.068–59.806)0.043Cluster 31 (0.8)Ref-Ref-Ref-Cluster 48 (3.6)4.837(0.598–39.117)0.1394.832(0.597–39.091)0.1404.767(0.589–38.582)0.143OR odds of ratio, CI confidence indexData were presented as number of case/total number (%), or OR (95% CI)1Adjusted by maternal age2Adjusted by adjusted for maternal age, pre-pregnancy BMI and GDMCluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) Table 3Association between different clusters and adverse pregnancy outcomesN (%)P valueCrude modelAdjusted model1Adjusted model2OR(95%CI)P valueOR(95%CI)P valueOR(95%CI)P valuePreeclampsia with severe features354 (31.5)0.216Cluster 1126 (33.4)1.619(1.026–2.555)0.0381.636(1.034–2.586)0.0351.645(1.04–2.601)0.033Cluster 2125 (31.8)1.505(0.954–2.372)0.0791.553(0.983–2.455)0.0591.532(0.969–2.423)0.068Cluster 331 (23.3)Ref-Ref-Ref-Cluster 472 (32.3)1.538(0.941–2.513)0.0861.588(0.969–2.601)0.0661.560(0.952–2.558)0.078Preterm birth117 (10.4)0.016Cluster 139 (10.3)3.663(1.283–10.459)0.0153.687(1.291–10.533)0.0153.701(1.295–10.573)0.015Cluster 251 (13.0)4.735(1.677–13.367)0.0034.853(1.717–13.714)0.0034.788(1.693–13.535)0.003Cluster 34 (3.1)Ref-Ref-Ref-Cluster 423 (10.3)3.651(1.234–10.804)0.0193.742(1.263–11.084)0.0173.680(1.242–10.906)0.019Low birth weight168 (14.9)< 0.001Cluster 150 (13.3)3.853(1.502–9.884)0.0053.871(1.508–9.932)0.0053.891(1.516–9.986)0.005Cluster 278 (19.8)6.240(2.468–15.776)< 0.0016.352(2.511–16.072)< 0.0016.271(2.478–15.870)< 0.001Cluster 35 (3.8)Ref-Ref-Ref-Cluster 435 (15.7)4.691(1.789–12.3)0.0024.773(1.819–12.523)0.0014.693(1.788–12.318)0.002Very low birth weight47 (4.2)0.079Cluster 115 (4.0)5.387(0.705–41.185)0.1055.386(0.704–41.177)0.1055.416(0.708–41.414)0.104Cluster 223 (5.9)8.081(1.081–60.436)0.0428.072(1.079–60.396)0.0427.991(1.068–59.806)0.043Cluster 31 (0.8)Ref-Ref-Ref-Cluster 48 (3.6)4.837(0.598–39.117)0.1394.832(0.597–39.091)0.1404.767(0.589–38.582)0.143OR odds of ratio, CI confidence indexData were presented as number of case/total number (%), or OR (95% CI)1Adjusted by maternal age2Adjusted by adjusted for maternal age, pre-pregnancy BMI and GDMCluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) Association between different clusters and adverse pregnancy outcomes OR odds of ratio, CI confidence index Data were presented as number of case/total number (%), or OR (95% CI) 1Adjusted by maternal age 2Adjusted by adjusted for maternal age, pre-pregnancy BMI and GDM Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS), Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS) Fig. 4Lipid profile patterns across four metabolic subtypes. A Standardized Z-scores of lipid parameters (TC, TG, LDL, HDL) by cluster. (i) LMS (Cluster 1): Uniformly low levels across all lipids; (ii) HHS (Cluster 2): Elevated HDL (+1.2 SD) with reduced TG (-0.8 SD); (iii) HTS (Cluster 3): High TG (+1.5 SD) and TC (+1.1 SD) but low LDL/HDL; (iv) HMS (Cluster 4): Pan-hyperlipidemia (all parameters > +0.9 SD). Data was presented as mean ± SEM. B Spearman correlation matrices of lipid profiles within each subtype. Significant correlations (p<0.01) are marked with asterisks. Notable findings include: strong negative HDL-TG correlation in HHS (r=-0.62); absent TC-HDL association in HHS (p=0.21). Abbreviations: TC, total cholesterol; TG, triglycerides; LDL, low-density lipoprotein; HDL, high-density lipoprotein; LMS, low metabolic subtype; HHS, high HDL subtype. Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS, Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS). Lipid profile patterns across four metabolic subtypes. A Standardized Z-scores of lipid parameters (TC, TG, LDL, HDL) by cluster. (i) LMS (Cluster 1): Uniformly low levels across all lipids; (ii) HHS (Cluster 2): Elevated HDL (+1.2 SD) with reduced TG (-0.8 SD); (iii) HTS (Cluster 3): High TG (+1.5 SD) and TC (+1.1 SD) but low LDL/HDL; (iv) HMS (Cluster 4): Pan-hyperlipidemia (all parameters > +0.9 SD). Data was presented as mean ± SEM. B Spearman correlation matrices of lipid profiles within each subtype. Significant correlations (p<0.01) are marked with asterisks. Notable findings include: strong negative HDL-TG correlation in HHS (r=-0.62); absent TC-HDL association in HHS (p=0.21). Abbreviations: TC, total cholesterol; TG, triglycerides; LDL, low-density lipoprotein; HDL, high-density lipoprotein; LMS, low metabolic subtype; HHS, high HDL subtype. Cluster 1 (low metabolic subtype, LMS), Cluster 2 (high HDL subype, HHS, Cluster 3 (high triglyceride subtype, HTS), and Cluster 4 (high metabolic subtype, HMS). Discussion Key findings This study is the first to apply unsupervised clustering to delineate maternal circulating lipid metabolic heterogeneity in preeclampsia, identifying four clinically distinct subtypes: (1) low metabolic subtype (LMS): global lipid reduction (30% of cohort); (2) high HDL subtype (HHS): elevated HDL (+ 1.2 SD) with reduced TG; (3) high TG subtype (HTS): high TG/TC but low LDL/HDL (paradoxically favorable outcomes); and (4) high metabolic subtype (HMS): pan-hyperlipidemia (highest BP/proteinuria). Crucially, HTS exhibited the lowest rates of preterm birth (3.1% vs. 13% in HHS) and compromised fetal growth (LBW 3.8% vs. 19.8%), despite elevated triglycerides. Mechanistic insights Our results suggested maternal circulating lipid phenotypes may reflect distinct placental adaptations or dysfunctions. The HTS phenotype mirrors previous findings where elevated cord blood TG coexisted with fetal growth restriction [24–26]. For instance, the HTS phenotype could represent an adaptive upregulation of placental triglyceride lipases (LPL) to enhance fatty acid delivery to the fetus under metabolic stress [25]. Larqué et al. reported that maternal hypertriglyceridemia in late pregnancy is associated with increased placental fatty acid transport capacity, possibly supporting fetal growth under metabolic stress [27]. The observed HDL paradox warrants further investigation. In HHS subtype, the elevated HDL in HHS might be dysfunctional due to pro‑inflammatory subspecies (e.g., Apo‑CIII‑enriched) impairing endothelial function [28]. Since HDL typically promotes cholesterol efflux and endothelial protection [26], its elevated levels in HHS failed to improve outcomes, possibly due to loss of anti-inflammatory capacity in preeclamptic HDL [28], and genetic variants (e.g., SCARB1 mutations) that alter HDL function [29, 30]. Future prospective, longitudinal studies would be required before any translational implications considered. In addition, although LMS and HMS differ primarily in lipid magnitude, they both show similar adverse outcomes clinical outcomes, suggesting they may represent ends of a continuous metabolic spectrum. Further studies with functional lipid assays are needed to determine whether these subtypes reflect distinct biological mechanisms. Clinical implications Third-trimester lipid profiling is potential for identifying high-risk subtypes (e.g., LMS with 33.4% and HMS with 32.3% severe features). The lipid subtypes identified in this study may reflect disease severity or compensatory responses. While these lipid subtypes may inform future risk stratification, these findings are hypothesis-generating. Whether simplified lipid thresholds applied for classify these phenotypes in real-world settings, requires further investigation. From a clinical perspective, by risk-stratification approaches, identified women with the LMS and HMS subtype who are at higher risk of severe features may benefit from closer monitoring. For example, recent evidence suggests that statins may be considered in selected high-risk cases of preeclampsia in the third trimester, particularly for women with hyperlipidemia who are at increased cardiovascular risk [31, 32]. Notably, mechanistic interpretation, such as a compensatory placental metabolic adaptation, requires direct experimental validation. The HTS subtype, characterized by elevated triglycerides and higher BMI, was unexpectedly associated with more favorable neonatal outcomes. Alternatively, it may indicate a compensatory placental metabolic adaptation. Notably, lower blood pressure and proteinuria in HTS suggest less severe placental dysfunction overall. Omega-3 fatty acid supplementation might be hypothesized to support placental triglyceride metabolism by serving as substrates for oxidation and reducing inflammatory lipid mediators, potentially offering a targeted approach [33, 34]. These interpretations remain speculative and require validation. Also, LMS subtype characterized by globally low lipid levels, was associated with worse fetal outcomes than the HTS subtype. It is possible that a state of relative lipid deficiency in the LMS cluster reflects a more severe form of placental dysfunction, where the placenta is inadequately acquire or synthesize lipids essential for fetal growth and steroid hormone production. Possible mechanism may be placental insufficiency in some cases of fetal growth restriction unrelated to preeclampsia. Importantly, because this is a retrospective observational study, we cannot rule out that the associations between lipid subtypes and outcomes are confounded by differences in clinical management. For example, patients in the HMS subtype had the highest blood pressure and proteinuria and likely received more aggressive antihypertensive therapy and earlier delivery, both of which independently affect neonatal outcomes. Conversely, the HTS subtype had the mildest disease and may have been managed more expectantly. Thus, the reported associations should not be interpreted as causal or as evidence that lipid profiles directly determine outcomes. Limitations This study acknowledges several limitations. First, lipid measurements were obtained in the third trimester, but the exact gestational week at sampling was not recorded. Considering that lipid levels may change exponentially across late pregnancy, differences in sampling timing could lead to the same patient being assigned to different lipid subtypes or could confound the observed associations between subtypes and outcomes. Although we could not adjust for gestational age at sampling retrospectively, future studies standardizing sampling timepoints are required to validate our findings. Second, the forced reliance on four conventional lipid markers, including TC, TG, LDL-C and HDL-C, may oversimplify the complexity of placental and maternal lipid metabolism. Unmeasured confounders (e.g., dietary lipids, pre-pregnancy dyslipidemia, medication use of aspirin and statins, treatment intensity, timing of diagnosis, or clinician management decisions) may influence outcomes. Environmental factors such as diet and physical activity, as well as genetic predispositions, which may influence the reliability of the outcomes. Residual confounding, such as clinical decision-making (e.g., timing of delivery, indication for preterm birth, antenatal corticosteroid administration, and severity-driven surveillance) may influence both maternal lipid measurements and neonatal outcomes. Despite the DAG-based adjustment is conceptually sound in our study, these factors are potentially difficult to fully capture in retrospective datasets. Therefore, the observed associations should not be interpreted as independent predictive effects of lipid subtypes beyond the adjusted variables. Third, despite that the four-cluster solution was statistically supported, the average silhouette width of 0.43 indicating moderate cluster cohesionsome inter-individual variability and overlap in lipid space particularly between LMS and HMS. Our identified subtypes represented data-driven phenotypes rather than discrete biological entities. External validation in an independent, prospective cohort is needed to confirm the stability and reproducibility of these four subtypes. Fourth, this study was conducted at a single tertiary referral center, which predominantly manages high-risk and complex preeclampsia cases. The prevalence of preeclampsia in our cohort is at the upper end of the reported global range, and the proportion of chronic hypertension was lower than expected in a real-world preeclampsia population. These factors suggest overrepresentation of more severe or atypical cases, which may limit generalizability to community or non-tertiary settings. The exclusion of 8.2% of eligible cases due to missing lipid data could also introduce bias, although baseline characteristics did not differ significantly between included and excluded women. Fifth, our study lacked a normotensive control group, preventing direct comparison with uncomplicated pregnancies. Therefore, the identified subtypes should be considered hypothesis-generating until validated in independent, multicenter cohort. Conclusion In summary, this study establishes lipid-driven subtypes of preeclampsia with distinct clinical trajectories, suggesting placental lipid handling may be a key modulator in disorder heterogeneity. This heterogeneity underscores the importance of lipid-based stratification for understanding placental dysfunction in preeclampsia. Supplementary Information Supplementary Material 1. Supplementary Material 1.