A stage-based dynamic model for predicting bronchopulmonary dysplasia in very preterm infants using longitudinal lung ultrasound: a retrospective cohort study

Background Bronchopulmonary dysplasia (BPD) remains a major complication in very preterm infants (VPIs), and static clinical models may not capture evolving neonatal lung injury. We assessed whether a stage-based model using serial lung ultrasound (LUS) could improve early BPD prediction. Methods In this retrospective single-center cohort, VPIs born at <32 weeks' gestation between January 2021 and December 2022 underwent LUS on postnatal days 7, 14, 21, and 28. Early clinical variables were collected, and BPD was defined by the 2018 NICHD criteria at 36 weeks’ postmenstrual age. Single-time-point models were compared with stage-based dynamic models incorporating prior predicted risk and later LUS data, using bootstrap out-of-bag validation. Results Among 269 VPIs, 174 developed BPD. LUS scores were higher in infants who developed BPD at all assessed time points. The best-performing single-time-point model used day-14 LUS (AUC 0.853, 95% CI 0.811–0.892). The three-stage random forest model achieved the highest point-estimate discrimination (AUC 0.874, 95% CI 0.835–0.909). Compared with the best-performing single-time-point model, the three-stage strategy showed an improvement in AUC of +0.021 (95% CI +0.005-+0.038). Compared with the two-stage strategy, the incremental difference in AUC was +0.009 (95% CI +0.001 to +0.018), whereas extension to the four-stage strategy showed no clear additional improvement (ΔAUC 0.006, 95% CI−0.004 to +0.016). In the three-stage model, prior-stage predicted risk and day-21 LUS remained independently associated with BPD. Conclusions Serial LUS-based risk updating provided incremental prognostic information beyond single-time-point assessment. The three-stage model incorporating information through postnatal day 21 achieved the highest point-estimate discrimination, whereas extension to day 28 provided no clear additional predictive gain.