Integration of Homocysteine-folate metabolic spectrum and machine learning for early prediction of preeclampsia

In addition, new biomarkers such as key substances of homocysteine-folate (Hcy-FA) metabolism spectrum are expected to provide parameter basis for machine learning models of PE early prediction. Objective: The purpose of this study is to develop, train and test an automated machine learning model for predicting the occurrence of preeclampsia. Furthermore, to verify the early predictive value of Hcy-FA metabolic profile biomarkers for PE.

Methods

We retrospectively collected 517 pregnant women who visited the Xuzhou Maternal and Child Health Hospital and the Affiliated Xuzhou Municipal Hospital of Xuzhou Medical University from November 2022 to July 2025. Clinical data and laboratory data were collected. Meanwhile, biomarkers such as homocysteine (HCY), S-adenosylhomocysteine (SAH), S-adenosylmethionine (SAM), 5-methyltetrahydrofolate (5-Methyl-THF), folic acid and 5-formyltetrahydrofolate (5-Formyl-THF) were detected by liquid chromatography-tandem mass spectrometry (LC-MS / MS) from the serum of pregnant women.

A total of 56 clinical features and 6 biomarkers were generated. After LASSO regularization screening and data cleaning, the remaining 29 core clinical features and 6 biomarkers were obtained. We trained four different models, random forest (RF), extreme random tree (ET), histogram gradient boosting tree (HGB) and logistic regression (LR).

The grid search method was used for hyperparameter training. All results were evaluated by a 5-fold stratified cross-validation scheme. Results: The comprehensive prediction ability of the three complex machine learning models ( RF, ET, HGB ) is better than that of the traditional LR model, with great performances, which indicating that the prediction ability of the nonlinear machine learning model is better than that of the traditional linear model.

The key substances in the Hcy-FA metabolic profile have also been shown to present certain predictive value as biomarkers, of which SAH has the greatest contribution. Conclusions: Compared with the current traditional clinical standard technology, the machine learning model based on Hcy-FA metabolic spectrum is an effective method to improve the prediction of high-risk pregnant women with PE.