Machine learning for the prediction of spontaneous preterm birth using early second and third trimester maternal blood gene expression: A cautionary tale

Spontaneous preterm birth (sPTB) remains a significant global health challenge and a leading cause of neonatal mortality and morbidity. Despite advancements in neonatal care, the prediction of sPTB remains elusive, in part due to complex etiologies and heterogeneous patient populations. This study aimed to validate and extend information on gene expression biomarkers previously described for predicting sPTB using maternal whole blood from the All Our Families pregnancy cohort study based in Calgary, Canada. The results of this study are two-fold: first, using additional replicates of maternal blood samples from the All Our Families cohort, we were unable to repeat the findings of a 2016 study which identified top maternal gene expression predictors for sPTB. Second, we conducted a secondary analysis of the original gene expression dataset from the 2016 study using five modelling approaches (random forest, elastic net regression, unregularized logistic regression, L2-regularized logistic regression, and multilayer perceptron neural network) followed by external validation using a pregnancy cohort based in Detroit, USA. The top performing model (random forest classification) suggested promising performance (area under the receiver operating curve, AUROC 0.99 in the training set), but performance was significantly degraded on the test set (AUROC 0.54) and further degraded in external validation (AUROC 0.50), suggesting poor generalizability, likely due to overfitting exacerbated by a low feature-to-noise ratio. Similar performance was observed in the other four learning models. Prediction was not improved when using higher complexity machine learning (e.g., neural network) approaches over traditional statistical learning (e.g., logistic regression). These findings underscore the challenges in translating biomarker discovery into clinically useful predictive models for sPTB. This study highlights the critical need for rigorous methodological safeguards and external validation in biomarker research. It also emphasizes the impact of data noise and overfitting on model performance, particularly in high-dimensional omics datasets. Future research should prioritize robust validation strategies and explore mechanistic insights to improve our understanding and prediction of sPTB.