Using Machine Learning to Expedite the Screening of Environmental Factors Associated with the Risk of Spontaneous Preterm Birth: From Exposure Mixtures to Key Molecular Events

Spontaneous preterm birth (SPB) is affected by various
environmental
exposures. However, there is still an urgent need to efficiently integrate
exposomic information to build its prediction model and unveil the
potential toxic pathways. Here, we conducted a nested case-control
study by recruiting 30 women with SPB delivery (cases) and 30 women
without (controls) at their early pregnancy. We analyzed various biomarkers
of external chemical exposure, lipidomics, and immunity, resulting
in 1081 exposure features. A logistic regression model (LR) was used
to screen potential risk factors, and four statistical learners were
used to establish SPB prediction models. Overall, the serum lipid
concentration in cases was higher than in controls, while this was
not the case for chemical and immune biomarkers. Random forest (RF)
and extreme gradient boosting (XGboost) models had a relatively higher
prediction accuracy of > 90%. Glycerophospholipids (GP) were the
most
abundant lipidomic features screened by LR, RF, and XGboost models,
followed by glycerolipids and sphingolipids, shown as well as by enrichment
analysis. Moreover, FA(21:0) had the largest contribution to the prediction
performance. Maternal exposure to various elements can contribute
to SPB risk due to their interaction with GP metabolism. Therefore,
it is promising to use exposomic data to predict SPB risk and screen
key molecular events.