Risk factors for prolonged respiratory support in late preterm infants: a LASSO-Cox regression analysis

Objective To investigate determinants of respiratory support duration and construct an interpretable weaning prediction model for late preterm infants. Methods A retrospective cohort study enrolled 365 late preterm infants (gestational age 34 + 0–36 + 6 weeks). Respiratory support duration (days) was the survival outcome, with successful weaning as the event. After Pearson correlation screening (|r| > 0.75 excluded), LASSO-Cox regression (10-fold cross-validation) selected predictors, followed by multivariable Cox modeling. Discrimination was assessed by C-index and time-dependent AUC at days 3, 5, and 8, with optimism-corrected bootstrap validation. Calibration was evaluated using calibration curves, Brier scores, and decision curve analysis. SHAP analysis quantified feature contributions on the log-hazard scale. Sensitivity analyses verified robustness. Results Multiple pregnancy (HR = 1.289, 95% CI: 1.038–1.601, P = 0.022), superoxide dismutase (SOD) (HR = 1.014, 95% CI: 1.009–1.019, P < 0.001), and albumin-to-globulin ratio (A/G) (HR = 1.130, 95% CI: 1.001–1.275, P = 0.049) were risk factors. Nasal continuous positive airway pressure (NCPAP) (HR = 0.703, 95% CI: 0.564–0.878, P = 0.002), lymphocyte percentage (LYM_PC) (HR = 0.987, 95% CI: 0.978–0.997, P = 0.008), and lactate dehydrogenase (LDH) (HR = 0.991, 95% CI: 0.984–0.998, P = 0.013, per 10-unit increase) were protective factors. Pulmonary surfactant showed no association ( P = 0.652). The model achieved a C-index of 0.677 (optimism-corrected 0.661) and td-AUC of 0.732, 0.769, and 0.782 at days 3, 5, and 8. Calibration was acceptable, DCA showed net benefit at day 5, and SHAP identified SOD, NCPAP, and LYM_PC as primary drivers. Conclusion The LASSO-Cox nomogram demonstrated moderate-to-good discrimination. Multiple pregnancy and elevated SOD are warning signs for prolonged oxygen dependency, while early NCPAP can shorten respiratory support. Nomogram visualization with SHAP interpretability provides a transparent basis for individualized weaning assessment.