Machine learning models for prediction of continuous positive airway pressure failure and high-flow nasal cannula failure in preterm neonates with respiratory distress.

This study aims to develop and internally validate separate machine learning models predicting continuous positive airway pressure (CPAP) failure and high-flow nasal cannula (HFNC) failure in preterm neonates. This retrospective cohort study included 411 preterm neonates started on non-invasive respiratory support between September 2023 and September 2025, 195 on CPAP and 216 on HFNC. CPAP failure was defined as escalation to invasive mechanical ventilation, HFNC failure as escalation to CPAP. Seven algorithms, LASSO logistic regression, random forest, XGBoost, LightGBM, support vector machine, naive Bayes, and k-nearest neighbors, were compared independently within each stratum using nine clinically selected predictors for CPAP and ten for HFNC, five-fold cross-validation, 1000-fold bootstrap confidence intervals, Platt calibration, decision curve analysis, and SHAP based explainability. Escalation occurred in 34 of 195 CPAP infants (17.4%) and 86 of 216 HFNC infants (39.8%). For CPAP failure, LASSO logistic regression is the most defensible model given the events-per-variable constraint (EPV 3.78), with AUROC 0.849 (95% CI 0.765-0.915). XGBoost showed the highest AUROC point estimate (0.876) but a larger apparent versus cross-validated gap. For HFNC failure, LASSO logistic regression achieved the highest discrimination (AUROC 0.773, 95% CI 0.662-0.837), with risk distributed across maternal chorioamnionitis, recurrent apnea, respiratory rate, and SpO2 trend. Both models showed positive net benefit over treat-all and treat-none strategies across a wide range of threshold probabilities. After Platt scaling, all calibration slopes fell between 0.924 and 1.027.

conclusionBedside respiratory, laboratory, and antenatal variables available within 6 h of starting non-invasive support may predict escalation risk in preterm neonates, on internal validation only. CPAP failure and HFNC failure differ in underlying structure and are best served by different predictors and different algorithms. External validation is needed before clinical implementation. WHAT IS KNOWN: • CPAP failure is linked to FiO2 requirement and RDS severity, with single-algorithm models predicting it in preterm neonates. • HFNC failure predictors come mainly from non-preterm populations, and no machine learning model for it exists in preterm neonates. WHAT IS NEW: • Seven algorithms were compared for CPAP and HFNC failure separately, using SHAP to show how each variable pushes an infant's risk up or down. • SHAP analysis showed CPAP failure risk concentrates around gas exchange status, while HFNC failure risk spreads across inflammation, apnea, and respiratory trend.