Integrating urine metabolomic biomarkers and machine learning algorithms to predict preeclampsia

Preeclampsia (PE) remains a leading cause of maternal and perinatal morbidity and mortality. This study aimed to identify urinary metabolites as potential biomarkers for predicting PE by integrating metabolomic profiling with machine learning algorithms.

The untargeted metabolomics of urine samples were performed in three cohorts: healthy pregnant women, gestational hypertension (GH), and PE patients. Differentially expressed metabolites were identified using multivariate statistical analyses. Subsequently, a predictive model for preeclampsia was developed and validated through four machine learning algorithms.

Metabolomic profiling identified 55 significantly dysregulated metabolites in PE compared to controls, while 22 metabolic signatures were observed between the GH and PE cohorts. Pathway analysis revealed vitamin B6 metabolism, steroid hormone biosynthesis, and histidine metabolism as core dysregulated pathways in PE pathogenesis. Next, LASSO regression selected four predictive metabolites: estriol-17 glucuronide, diethylphosphate, 4-deoxythreonic acid, and taurine. Notably, estriol-17 glucuronide demonstrated superior predictive accuracy compared to other metabolites. Furthermore, a machine learning model incorporating these four metabolic biomarkers was constructed for PE prediction. The XGBoost model showed significantly better prediction efficacy (94% accuracy, AUC = 0.976) compared to other machine learning models. In addition, estriol-17-glucuronide was significantly positively correlated with blood pressure.

This study identified four important urinary biomarkers and constructed an XGBoost-based predictive model for PE early detection. These findings provide a noninvasive approach for early PE screening in clinic and insights into its metabolic pathophysiology.