Comparative analysis of machine learning techniques in metabolomic-based preterm birth prediction

Machine learning (ML), with advancements in algorithms and computations, is seeing an increased presence in life science research. This study investigated several ML models' efficacy in predicting preterm birth using untargeted metabolomics from serum collected during the third trimester of gestation.

Samples from 48 preterm and 102 term delivery mothers from the All Our Families Cohort (Calgary, AB) were examined. Four ML algorithms: Partial Least Squares Discriminant Analysis (PLS-DA), linear logistic regression, artificial neural networks (ANN), Extreme Gradient Boosting (XGBoost) - with and without bootstrap resampling were used to examine the small-scale clinical dataset for both model performance and metabolite interpretation.

Model performance was evaluated based on confusion matrices, area under the receiver operating characteristic (AUROC) curve analysis, and feature importance rankings. Linear models such as PLS-DA and logistic regression demonstrated moderate classification performance (AUROC ≈ 0.60), whereas non-linear approaches, including ANN and XGBoost, exhibited marginal improvements. Among all models, XGBoost combined with bootstrap resampling achieved the highest performance, yielding an AUROC of 0.85 (95 % CI: 0.57-0.99, p < 0.001), indicating a significant improvement in classification accuracy. Metabolite importance, derived from Shapley Additive Explanations (SHAP), consistently identified acylcarnitines and amino acid derivatives as principal discriminative features. Pathway analysis revealed disruptions to tyrosine metabolism as well as phenylalanine, tyrosine and tryptophan biosynthesis to be associated with preterm delivery.

Our results highlight the complexity of metabolomics-based modelling for preterm birth and support an iterative, model-driven approach for optimizing predictive accuracy in small-scale clinical datasets.