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

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Reproductive Health

Background:

• Preterm birth (before 37 weeks) is a major public health concern with significant burdens.
• Early identification of preterm birth is crucial for intervention and improved outcomes.
• Preterm births are categorized into spontaneous and indicated subtypes, requiring distinct predictive approaches.

Purpose of the Study:

• To investigate the predictive ability of machine learning models for early identification of preterm birth in nulliparous women.
• To develop and compare targeted predictive models for spontaneous and indicated preterm births.
• To evaluate the efficacy of incorporating 'privileged information' into machine learning models for enhanced preterm birth prediction.

Main Methods:

• Analysis of the Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-be (nu-MoM2b) cohort.
• Development of an enhanced XGBoost model (XGBoost+) utilizing privileged information (adverse pregnancy outcomes, post-delivery physiology, maternal outcomes).
• Distinction between spontaneous and indicated preterm births for subtype-specific model development and evaluation using AUC.

Main Results:

• XGBoost-based models outperformed other traditional machine learning approaches (decision tree, random forest, logistic regression, SVM).
• The XGBoost+ model achieved an overall AUC of 0.72.
• XGBoost+ showed significant improvement for indicated preterm birth prediction (AUC 0.78 vs. 0.74 for XGBoost) but similar performance for spontaneous preterm birth (AUC 0.68 vs. 0.67).

Conclusions:

• Preterm birth is multifactorial, with distinct risk factors for spontaneous and indicated subtypes.
• The XGBoost+ model demonstrates strong predictive performance for indicated preterm birth, likely due to its ability to capture hypertension and preeclampsia-related factors.
• While indicated preterm birth is predictable with clinical data, spontaneous preterm birth prediction remains challenging, suggesting a need for proximal biological data in future research.