A machine learning-based risk prediction model for early preterm birth: development and prospective validation

Introduction: Preterm birth (PTB) remains a leading cause of neonatal morbidity and mortality worldwide, yet accurate and clinically applicable risk prediction tools during mid-pregnancy are still limited. Methods: We developed a machine learning-based model to predict PTB risk using routinely collected clinical and laboratory data obtained during mid-pregnancy. In a derivation cohort, candidate predictors were selected using least absolute shrinkage and selection operator (LASSO) regularization, followed by training and comparison of multiple machine learning algorithms. A light gradient boosting machine (LightGBM) model was selected as the final model. Model performance was evaluated in a temporally independent prospective validation cohort from the same tertiary maternity center in terms of discrimination, calibration, and clinical utility. A screening-oriented risk stratification strategy was applied based on the distribution of predicted probabilities. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) to quantify feature contributions. Results: In the internal test set of the derivation cohort, LightGBM demonstrated the best discriminative performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.842. In the prospective validation cohort, the model maintained stable discrimination and good calibration. Using a screening-oriented risk stratification approach, women classified as high risk exhibited a substantially higher incidence of PTB compared with the intermediate- and low-risk groups (31.8% vs. 6.0% and 2.5%, respectively), with clear separation of cumulative PTB incidence across gestational age. Decision curve analysis demonstrated a consistent net benefit across a range of low threshold probabilities. SHAP analysis revealed that inflammatory markers, liver-related biochemical indices, pregnancy complications, and nutritional indicators were key contributors to model predictions, supporting clinical interpretability. Conclusions: We developed a machine learning-based model for mid-pregnancy PTB risk prediction and evaluated it in a temporally independent prospective validation cohort from the same tertiary maternity center. The model demonstrated stable discrimination, good calibration, meaningful risk stratification, and transparent interpretability, supporting its potential use as a screening-oriented decision-support tool for targeted antenatal surveillance.