Multidimensional predictors of preterm birth risk among black and white primiparous women in the U.S.: insights from machine learning

Unmeasured contextual factors contribute to Black-White disparities in preterm birth (PTB), but their effects are difficult to isolate due to complex relationships with individual factors connected in non-linear ways. To address this, we applied explainable machine learning to model interactions between individual and contextual factors to predict PTB and identify its key predictors among non-Hispanic Black (NHB) and non-Hispanic White (NHW) primiparous women in the U.S. Elastic Net, Random Forest, and XGBoost models were developed using Pregnancy Risk Assessment Monitoring System and the Social Vulnerability Index data from nine U.S. states. SHAP (SHapley Additive exPlanations) values were computed to assess feature importance. Model performance was evaluated using the area under the ROC curve (AUC). Our models predicted PTB with high accuracy (AUC: 0.87–0.93) for NHB and NHW primiparous women, identifying both shared and distinct multidimensional predictors. Shared individual predictors included ≥ 9 prenatal care visits (protective; mean |SHAP| 0.42–1.58), adequate + prenatal care (risk-increasing; mean |SHAP| 0.69 for NHB and 1.18 for NHW), and gestational hypertension (risk-increasing; mean |SHAP| 0.17 and 0.20, respectively). Contextual socioeconomic status and household composition also contributed significantly to PTB prediction, with a stronger impact among NHB women. Explainble machine learning with SHAP values can accurately quantify the contribution of individual and contextual factors to PTB risk specific to NHB and NHW primiparous women. By integrating feature importance with the prevalence of risk factors, this approach offers actionable insights to identify priority areas for intervention and inform tailored preventive strategies aimed at reducing Black-White disparities in PTB.