Maternal Metabolism and Pregnancy: Predicting Pregnancy Outcomes Using Metabolomic Assessment
Date
2024-10-31
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Despite the significant health risks preterm birth poses to both mothers and babies, assessing the risk remains challenging due to its heterogeneous nature. Common risk factors, such as personal health history or lifestyle habits, have had some success in identifying at-risk mothers. However, their effectiveness is limited in asymptomatic and niche populations that lack distinct, identifiable preterm traits. Recent efforts have shifted toward exploring metabolic biomarkers for preterm prediction, aiming to capture the metabolic changes occurring during gestation. Nevertheless, most targeted compounds in these studies have been chosen based on their association with other underlying conditions, such as obesity or uterine disorders. With advancements in computational methods, particularly in machine learning, prediction models based on high-dimensional data have emerged as a promising new approach. This project aims to explore metabolic signatures associated with preterm birth by leveraging machine learning and untargeted metabolomics. We analyzed third-trimester serum samples from primigravid participants—first-time mothers—enrolled in the All Our Families (AOF) Cohort. Significant reductions in acylcarnitines and amino acid derivatives were identified, with various acylcarnitines, particularly butenylcarnitine, notably reduced in preterm mothers. These metabolites proved effective in predicting preterm birth, as shown by receiver operating characteristic (ROC) analysis. Next, we compared the performance of six different machine learning models in predicting preterm birth across a broader population from the AOF cohort, using Shapley Additive Explanations (SHAP) analysis to evaluate feature importance. We also assessed the impact of resampling techniques, such as bootstrapping. Linear models, including PLS-DA and linear logistic regression, demonstrated moderate predictive performance, while non-linear models like XGBoost and artificial neural networks (ANN) showed a slight advantage. Bootstrapping improved model accuracy and predictive strength, with varying degrees of enhancement across different models. Among the models tested, the bootstrap-resampled XGBoost model was the top performer in predicting preterm birth. SHAP analysis consistently identified acylcarnitines as the most significant class of metabolites in preterm prediction, with kynurenic acid also emerging as a contributing metabolite in several models following bootstrapping. The observed alterations in acylcarnitines suggest possible disruptions in lipid transport and energy metabolism in preterm mothers. The modeling results underscored the complexity of preterm prediction, while resampling techniques proved effective in mitigating the overtraining and low-variance challenges posed by the small sample size. This project introduces a novel approach to predicting preterm birth based on a combination of metabolites. As preterm birth continues to present significant risks to fetal health, the development of newer and more effective prediction models could improve maternal and fetal outcomes.
Description
Keywords
Preterm Birth, Gestation, Metabolomic, Machine Learning
Citation
Han, Y. C. (2024). Maternal metabolism and pregnancy: predicting pregnancy outcomes using metabolomic assessment (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca.