Improved Prediction of Preterm Birth Combining Machine Learning and Multi-Trait Polygenic Scores

Spontaneous preterm birth (sPTB) is a leading driver of neonatal morbidity and mortality, yet clinicians still lack accurate, scalable tools to identify high‑risk pregnancies early enough to target preventive care. Current prediction approaches rely on a limited set of clinical indicators (e.g., prior sPTB, midtrimester cervical length) and perform poorly for primiparous patients and those with incomplete obstetric histories—groups that include many individuals experiencing disparities in access to care. This R21 will develop and externally validate a clinically actionable prediction framework that integrates routinely collected prenatal clinical information with a single, scalable genetic assessment: a multi‑trait polygenic score (mPGS) capturing inherited liability across >100 traits relevant to sPTB pathways. Aim 1 will train and optimize prediction models in the Wayne State University (WSU) pregnancy cohort (~5,000 Black/African American participants; ~4,400 with high‑quality genotypes). We will compute trait‑specific polygenic scores using externally derived, well‑powered GWAS weights and combine them using PRSmix/PRSmix+, an elastic‑net mixing approach that yields a robust composite genetic predictor while penalizing uninformative components. This strategy is particularly well suited for clinical translation because it mitigates instability from any single polygenic score, improves portability in African‑ancestry populations by learning target‑cohort–specific mixing weights, and can leverage an inclusive library of available scores rather than relying only on ancestry‑matched resources. Models will be constructed in time‑anchored clinical windows (e.g., first‑trimester prediction), enabling evaluation of how early risk stratification can be achieved in real prenatal workflows. Aim 2 will generate new low‑pass sequencing data in the independent MOMS‑PI cohort and validate locked WSU models without retraining, providing a realistic estimate of performance in a separate clinical setting. We will emphasize calibration and “calibration transfer,” reporting minimal recalibration procedures that mirror how risk tools are implemented across health systems. Deliverables will include: (i) an externally validated mPGS‑enhanced risk model optimized for African American women, (ii) quantified improvement over covariate‑only and best single‑score baselines, and (iii) an implementation workflow suitable for extension to other pregnancy cohorts and future prospective clinical evaluation. If successful, this project will provide a practical foundation for earlier identification of high‑risk pregnancies and more equitable deployment of precision prenatal prevention strategies.