Predicting 2-year neurodevelopmental outcomes in extremely preterm infants using graphical network and machine...

Area of Science:

• Neonatal Medicine
• Machine Learning in Healthcare
• Neurodevelopmental Pediatrics

Background:

• Infants born extremely preterm (<28 weeks' gestation) face a high risk of neurodevelopmental impairment (NDI).
• Approximately 50% of survivors exhibit moderate to severe NDI by 2 years of age.
• Accurate prediction models are crucial for early intervention and improved outcomes.

Purpose of the Study:

• To develop and compare novel predictive models for neurodevelopmental outcomes in extremely preterm infants.
• To evaluate the efficacy of machine learning, specifically Bayesian Additive Regression Trees (BART), in predicting NDI and Bayley Scales of Infant and Toddler Development (Bayley) scores.
• To assess the contribution of baseline characteristics, clinical care, and environmental exposures to predictive accuracy.

Main Methods:

• Utilized a prospective database of 692 infants from the Preterm Epo Neuroprotection (PENUT) Trial.
• Developed three BART models: 1) using 5 NICHD Extremely Preterm Birth Outcomes Tool variables, 2) using 21 clinical variables, and 3) a hypothesis-free approach with 133 variables.
• Predicted NDI and continuous Bayley subscale scores at 2-year follow-up.

Main Results:

• The 5-variable NICHD model explained 3-4% of variance (AUROC 0.62).
• The 21-variable model improved prediction (12-20% variance, AUROC 0.77).
• The hypothesis-free BART model achieved the highest accuracy (20-31% variance, AUROC 0.87 for severe NDI), with higher transfusion volume being a key predictor.

Conclusions:

• Machine learning (BART) significantly enhanced predictive accuracy for NDI and Bayley scores compared to existing tools.
• Despite improvements, average prediction errors for Bayley scores remain substantial (approx. 1 SD), indicating limitations in precise outcome prediction.
• The findings underscore the ongoing need for comprehensive long-term follow-up for all extremely preterm infants due to the inherent variability in neurodevelopmental outcomes.