Parental health at preconception and gestational age at birth: evidence from a population-based cohort using double machine learning

Abstract

Background: Gestational age at birth is shaped by complex maternal and paternal factors prior to conception, yet few studies have integrated multidimensional clinical indicators from both parents to estimate their quantitative effects.

Methods: Using a large retrospective cohort with routinely collected preconception health data from both mothers and fathers, we applied a double machine learning (DML) framework combining Lasso, Random Forest and XGBoost regressors to estimate associations between parental biomarkers and gestational age at birth. Model interpretability was enhanced through Shapley Additive Explanations (SHAP) analysis, stratified interaction testing and estimation of individualized treatment effects (ITE).

Results: Elevated maternal fasting glucose, alanine aminotransferase, platelet count and anti-hepatitis B core seropositivity were consistently associated with shortened gestational age at birth across all DML models. For instance, a 10 mmol/L increase in fasting glucose corresponded to 2.8-4.9 days shorter gestation. Paternal Treponema pallidum seropositivity and increased monocyte proportion, defined as the fraction of monocytes within total circulating white blood cells, also demonstrated significant associations, with the former linked to gestational shortening and the latter to modest extension. Stratified and interaction analyses revealed that paternal immune markers modified the associations of maternal metabolic indicators with gestational age at birth. SHAP-based interpretation confirmed model consistency and ITE analysis indicated marked heterogeneity, particularly for fasting glucose.

Conclusions: This study demonstrates the utility of interpretable DML methods for quantifying the effects of multidimensional preconception health indicators from both parents on gestational age at birth. Our findings support integrating maternal and paternal screening into preconception risk assessments to enable early, targeted prevention of preterm birth. Given the established links between preterm birth and neonatal surgical conditions, early risk identification via parental screening may help inform perinatal care strategies and optimize resource allocation in pediatric surgery.

Keywords: Child Health; Health Services; Information Technology.

Figure 1. Study population flowchart.

Figure 2. SHAP summary plots identifying top predictors of gestational age at birth across three DML models. (1) SHAP values represent the contribution of each feature to the model’s prediction for individual samples. The absolute magnitude of SHAP values reflects the strength of the feature’s influence on the prediction. (2) Y-axis lists the top 10 features ranked by their overall importance, measured by the sum of absolute SHAP values across all samples. Higher positions indicate stronger overall contributions. X axis shows the SHAP value distribution across all samples for each feature, reflecting both the direction and magnitude of its effect. Positive SHAP values indicate a feature increases the predicted gestational age, while negative values indicate a decrease. (3) Each dot represents one sample. The color gradient indicates the actual feature value: red for high values or positive test results, blue for low values or negative results. For example, in the rightmost panel (Random Forest-DML), higher maternal ALT values (red dots) are associated with negative SHAP values, indicating a shorter predicted gestational duration. ALT, alanine aminotransferase; DML, double machine learning; SHAP, Shapley Additive Explanations.

Figure 3. Stratified associations of maternal preconception indicators with gestational days at birth across DML models, by paternal inflammatory status. Effect estimates (days, 95% CI) represent the change in gestational age at birth (in days) per specified unit increase in each continuous maternal indicator. Fasting glucose and ALT were modeled per 10 mmol/L and 100 U/L increase, respectively; platelet count per 500 × 10⁹/L increase. Models were adjusted for maternal and paternal demographic and clinical covariates. P values are derived from three model-based estimations (Lasso-DML, Random Forest-DML and XGBoost-DML). ALT, alanine aminotransferase; DML, double machine learning; RF, random forest; XGBoost, extreme gradient boosting; CI, confidence interval.