Predicting Neonatal Respiratory Outcomes Using Machine Learning: A Systematic Review Through a Life Course Lens.

BACKGROUND

Neonatal respiratory outcomes remain leading drivers of neonatal intensive care unit (NICU) morbidity, mortality, and prolonged hospitalization, underscoring the need for accurate early risk prediction through artificial intelligence (AI) and machine learning (ML) approaches.

PURPOSE: To conduct a Preferred Reporting Items for Systematic Reviews and Meta-Analyses-guided systematic review synthesizing AI/ML models predicting neonatal respiratory outcomes through a Life Course Health Development (LCHD) lens, examining whether models incorporate developmental timing, cumulative processes, and contextual factors.

DATA SOURCES: PubMed, Embase, and Web of Science were searched from January 1, 2014, to July 18, 2025.

STUDY SELECTION: Included peer-reviewed studies developing or validating AI/ML prediction models for neonatal respiratory outcomes with performance metrics. Excluded non-ML regression, non-original research, and non-English studies. Of 4319 records initially identified, 1780 unique records were screened, 33 underwent full-text review (0.76%); 16 met selection criteria (0.37%).

DATA EXTRACTION: Two reviewers independently extracted study characteristics, inputs, algorithms, validation strategies, and performance metrics. Risk of bias was assessed using PROBAST + AI. Disagreements were resolved by consensus.

RESULTS

Studies addressed bronchopulmonary dysplasia (n = 8), respiratory distress syndrome (n = 4), apnea of prematurity (n = 3), and massive pulmonary hemorrhage (n = 1). Discrimination was strong in internal validation (several models with an area under the receiver operating characteristic curve ≥0.85). External validation was rare (2/16); calibration reporting was sparse. No studies predicted postdischarge respiratory outcomes or included comprehensive determinants of health.

IMPLICATIONS FOR PRACTICE AND RESEARCH: Although ML models show promise for NICU risk stratification, responsible clinical adoption requires multisite external validation, routine calibration, and nursing-informed designs incorporating development context. See Video Abstract, Supplemental Digital Content.