Metabolic trajectories in twin pregnancies: longitudinal lipid and glucose profiling improves prediction of preterm birth

Abstract Background Preterm birth (PTB) complicates nearly half of all twin pregnancies, yet current risk-stratification tools rely heavily on static anatomical markers such as cervical length. We hypothesised that longitudinal metabolic profiling across gestation could refine early identification of women at highest risk. Methods We conducted a retrospective cohort study including 601 dichorionic and monochorionic twin pregnancies delivered at Fujian Provincial Maternity and Children’s Hospital (January 2023–December 2024). Fasting glucose and six lipid parameters were measured in early (< 14 weeks), middle (14–27⁺⁶ weeks), and late (≥ 28 weeks) gestation. Group-based trajectory modelling identified latent metabolic phenotypes. Using partitioning around medoids clustering, we integrated glucose/triglyceride trajectories, mid-trimester cervical length, chorionicity, and neutrophil-to-lymphocyte ratio into four risk phenotypes. Multivariable logistic regression evaluated associations with PTB < 37 weeks. A Least Absolute Shrinkage and Selection Operator(LASSO)-penalised model incorporating continuous metabolic measures and trajectory classes was constructed and internally validated using five-fold cross-validation and bootstrap resampling. Model performance was compared against a baseline clinical model and a phenotype-only model. Results The overall PTB rate was 49.1%. Seven metabolic indices yielded stable trajectories. A sustained hypercholesterolaemic trajectory (TC1) carried the highest PTB risk (96.2%), whereas a physiologically rising High-Density Lipoprotein Cholesterol(HDL-C) trajectory (HDL2) was associated with the lowest (21.7%). Four distinct phenotypes emerged: P1 (low risk, PTB 25.9%), P2 (intermediate, 38.6%), P3 (high risk, 76.5%), and P4 (very high risk, 79.8%). Compared with P1, adjusted odds ratios for PTB were 1.97 (95% CI 1.13–3.44), 6.90 (2.81–16.94), and 12.65 (6.67–23.98) for P2–P4, respectively. The LASSO-enhanced model (M2) achieved a cross-validated Area Under the Receiver Operating Characteristic Curve(AUC) of 0.902 (Brier score 0.126), outperforming both the clinical model (AUC 0.565) and the phenotype model (AUC 0.730). Key predictors included mid-trimester fasting glucose, TC, HDL-C, Low-Density Lipoprotein Cholesterol(LDL-C), apolipoprotein B(ApoB), and TC trajectory class. Sensitivity analyses, including multiple imputation and exclusion of iatrogenic PTB, confirmed model robustness. Conclusions Longitudinal metabolic trajectories delineate dynamic metabolic maladaptation and serve as a valuable complement to conventional clinical variables for PTB risk assessment. While our findings support the potential utility of this framework, it remains an exploratory tool that requires prospective validation before it can be considered a standard component of personalised antenatal surveillance.