Unsupervised Machine Learning for the Identification of Latent First-Trimester Obstetric Phenotypes Associated with...

Background/Objectives: Artificial intelligence may improve obstetric risk stratification by identifying clinically meaningful patterns that are not captured by single-outcome prediction models. This study evaluated whether unsupervised machine learning applied to first-trimester maternal, biophysical, and biochemical data could identify phenotypes associated with adverse obstetric outcomes. Methods: We analyzed 1583 first-trimester records that were ambispectively collected including maternal characteristics, obstetric history, mean arterial pressure, uterine artery pulsatility index, nuchal translucency, and PAPP-A. Principal component analysis followed by k-means clustering was used to derive AI phenotypes. Associations were tested for hypertensive, metabolic, delivery, neonatal, and actionable obstetric outcomes. Robustness of the clustering solution and phenotype-outcome associations was assessed using inverse probability weighting, false discovery rate correction, adverse-event burden scoring, cross-algorithm agreement analyses, and multiple sensitivity analyses. Results: Four clinically interpretable phenotypes were identified. Phenotype 4 represented a cardiometabolic-vascular subgroup that showed the greatest enrichment for adverse obstetric outcomes. It had higher rates of maternal metabolic or hypertensive complications (52.4%; RR 4.31) and clinically actionable adverse pregnancy outcomes (57.1%; RR 2.31). These associations remained significant after false discovery rate correction and were consistent in inverse probability weighting analyses. Phenotype 4 also had the highest cumulative adverse-event burden score (1.29 versus 0.35-0.47). Exploratory analyses demonstrated high specificity (97.2%) but limited sensitivity for clinically actionable adverse pregnancy outcomes. Incremental discrimination over conventional predictors was observed only for selected secondary outcomes, including early preterm birth and low Apgar score. Conclusions: Unsupervised AI phenotyping identified a cardiometabolic-vascular pregnancy subgroup associated with increased obstetric morbidity.