Maternal and Neonatal Determinants of Respiratory Outcome Following Second-Trimester PPROM: A Multi-Domain Machine Learning Analysis

Background: Preterm premature rupture of membranes (PPROM) before 32 weeks of gestation with prolonged latency is associated with substantial neonatal morbidity, including Dry Lung Syndrome (DLS), pulmonary hypoplasia (PH), bronchopulmonary dysplasia (BPD), and death. Accurate individualized risk stratification remains elusive, as the interacting contributions of amniotic fluid dynamics, inflammatory status, and microbiological burden are inadequately captured by traditional statistical approaches. Methods: We performed a retrospective, exploratory–predictive analysis of 66 pregnancies complicated by second-trimester PPROM with latency exceeding 14 days. Elastic Net and Random Forest models were trained across six clinically defined predictor domains using a multi-stage block modelling strategy. To address the clinically relevant distinction between antenatal and postnatal information, results are reported separately for Model A—comprising exclusively antenatal predictors available during expectant management (gestational age at PPROM, latency, amniotic fluid trajectory, inflammatory status, vaginal microbiome at admission)—and Model B, which additionally incorporates postnatal variables and characterizes the full mechanistic perinatal risk trajectory. Binary and ordinal outcomes included DLS, PH, BPD, intraventricular hemorrhage (IVH), and neonatal death. Pairwise interaction models were additionally computed to identify cross-domain risk constellations. Results: Distinct predictor architectures emerged per outcome. Pulmonary hypoplasia was most strongly associated with temporal features of oligohydramnios—particularly the persistence and timing of SDP < 1 cm—rather than isolated measurements. For DLS, the antenatal model (Model A) achieved AUC 0.776, driven by gestational maturity and inflammatory status; surfactant administration—a postnatal variable reflecting therapeutic response rather than an antenatal risk factor—dominated only the mechanistic Model B. Neonatal death was driven by a combined profile of respiratory support burden, amniotic fluid persistence, and co-morbidity. IVH showed consistently high ordinal predictability (accuracy 0.863), with amniotic fluid dynamics and microbiological burden as leading contributors. BPD remained the least linearly separable endpoint across all configurations. Conclusions: Multi-domain machine learning reveals outcome-specific, cross-domain risk architectures following second-trimester PPROM that are invisible to conventional statistical models. Longitudinal amniotic fluid trajectory is the dominant antenatal determinant of structural pulmonary morbidity, while microbiological burden independently shapes neurological risk. These findings support prospective validation of integrated ML-based risk stratification tools for individualized antenatal counselling in this high-risk population.