Differential Predictability of Preterm Birth Types: Strong Signals for Indicated Cases versus Limited Success in Spontaneous Preterm Birth.

Background: Preterm birth, defined as birth occurring before 37 weeks of gestation, poses a significant and enduring public health challenge, with substantial emotional and financial burdens on families and society. To identify preterm births early in pregnancy, we investigated the predictive ability of machine learning models in a nulliparous (first-time pregnancy) study cohort. Preterm births are categorized into two major types: indicated preterm birth, which occurs due to medical conditions such as preeclampsia or other maternal/fetal complications requiring early delivery, and spontaneous preterm birth, which involves the natural onset of preterm labor. Our research aims to develop predictive tools that could enable earlier intervention and improved outcomes for these vulnerable pregnancies.

Methods: Our study analyzed the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-be cohort (nu-MoM2b), comprising data from eight clinical sites throughout the United States from October 2010 to May 2014, including treatment, psychological, physiological, medical history, demographic, ultrasound, activity, toxicology, family history, pre-pregnancy diet, and genetic race. We distinguished between spontaneous and indicated preterm births to develop targeted predictive models for each subtype. We also used a novel approach to predict preterm birth called learning with privileged information, information available during training but often inaccessible during evaluation. Specifically, the set of privileged information that we utilized for PTB prediction includes the occurrence of adverse pregnancy outcomes (APOs), after-delivery physiology information, and maternal outcomes. We developed an enhanced model, XGBoost+, which incorporates this privileged information to improve predictive performance compared to traditional machine learning approaches.

Results: We selected XGBoost as our base model due to its robust performance with tabular data and its ensemble approach that effectively mitigates overfitting while capturing complex relationships between clinical variables, making it particularly well-suited for the heterogeneous risk factors associated with preterm birth prediction. XGboost-based models achieved higher AUC against all other models, including decision tree, random forest, logistic regression, and SVM for all visits. Our XGboost+ model, utilizing privileged information, achieved an AUC of 0.72. Analyzing the subcategories of preterm birth, XGboost+ achieved similar performance with XGboost for spontaneous preterm birth (0.68 AUC versus 0.67 AUC), but improvements were more significant for indicated preterm birth (0.78 versus 0.74). These results demonstrate the benefits of how information that is not typically utilized in traditional machine learning models can help build better models.

Conclusion: Our extensive analysis of this comprehensive set of risk factors revealed preterm birth as a multifaceted issue, with different risk factors associated with two subcategories of preterm birth - spontaneous and indicated. No-tably, we achieved significant success in predicting indicated preterm birth, demonstrating strong predictive performance (AUC 0.78) using our XGBoost+ model. This finding represents an important advancement, as indicated preterm birth is influenced mainly by conditions related to hypertension and preeclampsia, which our model effectively captured. While spontaneous preterm birth remains challenging to predict with clinical data alone, especially in early pregnancy, our research successfully differentiates between these subtypes and provides a valuable predictive tool for indicated preterm birth. The complexity of spontaneous preterm birth suggests that future research should focus on gathering more proximal biological data, including vaginal microbiota or raw cervical images, to complement our successful approach for indicated preterm birth prediction.