Skip to Content
BiomedicinesBiomedicines
  • This is an early access version, the complete PDF, HTML, and XML versions will be available soon.
  • Article
  • Open Access

27 September 2026

27 Pages

Exploratory Integration of Metabolomics and Machine Learning Identifies Candidate Signatures of Spontaneous Preterm Birth

,
,
,
and
1
School of Medicine, Universidad Industrial de Santander, Bucaramanga 680002, Colombia
2
MetCore—Metabolomics Core Facility, Vice-Presidency for Research, Universidad de los Andes, Bogota 111711, Colombia
3
School of Chemistry, Universidad Industrial de Santander, Bucaramanga 680002, Colombia
*
Author to whom correspondence should be addressed.

Abstract

Background/Objectives: Spontaneous preterm birth (sPTB) remains a major cause of neonatal morbidity and mortality. This study aimed to characterize first-trimester plasma metabolic alterations associated with sPTB and evaluate their discriminatory potential using metabolomics and machine learning. Methods: This retrospective case–control study included 199 pregnant women (64 sPTB and 135 term births) from the Colombian Cohort for the Early Prediction of Preterm Birth (COLPRET). First-trimester plasma was analyzed by Liquid Chromatography–Quadrupole Time-of-Flight Mass Spectrometry (LC-QTOF-MS). Multivariate and univariate analyses were used to characterize metabolic differences. Thirty candidate metabolites were selected through outcome-associated univariate analyses performed on the complete cohort before subsequent machine-learning model evaluation. Six machine-learning algorithms were evaluated using clinical, metabolomic, and integrated predictor sets within a nested stratified cross-validation framework. Results: Lipid-related metabolites, particularly lysophospholipids and phospholipids, predominated among the molecular alterations associated with sPTB. Lysophosphatidylcholine O-14:0 (LPC O-14:0) showed the highest individual discrimination with an area under the receiver operating characteristic curve (ROC-AUC) (AUC = 0.705). Models based on clinical data showed limited discrimination (out-of-fold [OOF] ROC–AUC: 0.500–0.555), whereas metabolomic and integrated models achieved OOF ROC–AUC values of 0.632–0.760 and 0.641–0.770, respectively. Metabolomic and integrated models generally showed higher ROC–AUC estimates than clinical models. However, adding the five clinical variables to the metabolomic predictor set did not significantly improve the ROC–AUC after correction for multiple comparisons (all false discovery rate [FDR]-adjusted q ≥ 0.060). Conclusions: First-trimester plasma metabolomics identified predominantly lipid-related alterations associated with subsequent sPTB. Metabolomic information showed greater discriminatory capacity than the five evaluated clinical variables, with no robust evidence of additional discrimination from their integration. These exploratory findings require validation in larger prospective cohorts before clinical utility can be established.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.