Abstract / Summary
Abstract Persistent pulmonary hypertension of the newborn (PPHN) remains a life-threatening disorder with considerable mortality despite the widespread use of inhaled nitric oxide (iNO). Current risk assessment relies on static indices and cannot capture the rapid physiological changes characteristic of this disease. We aimed to evaluate the association between dynamic oxygenation indices and mortality, and to develop a nonlinear joint modeling framework for individualized prediction in neonates receiving iNO therapy. In this retrospective cohort of 377 PPHN neonates, serial blood gas and ventilator data were used to compute oxygen saturation index (OSI) and oxygen index (OI). Linear and nonlinear joint models linking longitudinal OSI/OI trajectories with time-to-death were constructed via Bayesian estimation. Results Non-survivors showed persistently higher OSI and OI trajectories than survivors (p < 0.001). Among six candidate models, the nonlinear OSI-based joint model achieved the best performance (lowest DIC/WAIC, highest LPML) and excellent discrimination (24-h AUC = 0.93, cross-validation AUC = 0.90, validation AUC = 0.84), substantially outperforming linear (validation AUC: 0.78–0.82), OI-based (validation AUC: 0.75–0.83), time-varying Cox (validation AUC: 0.74–0.79), and nSOFA-based models (validation AUC: 0.68–0.70). Dynamic predictions demonstrated that early OSI trajectories strongly influenced survival probability. Dynamic OSI trajectories are independent predictors of mortality in PPHN neonates. Nonlinear OSI-based joint modeling enables individualized, time-updated mortality prediction within the first 24 h of iNO therapy and may inform early clinical decision-making.