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

Background: Infants born extremely preterm (<28 weeks’ gestation) are at high risk of neurodevelopmental impairment (NDI) with 50% of survivors showing moderate or severe NDI when at 2 years of age. We sought to develop novel models by which to predict neurodevelopmental outcomes, hypothesizing that combining baseline characteristics at birth with medical care and environmental exposures would produce the most accurate model. Methods: Using a prospective database of 692 infants from the Preterm Epo Neuroprotection (PENUT) Trial, which was carried out between December 2013 and September 2016, we developed three predictive algorithms of increasing complexity using a Bayesian Additive Regression Trees (BART) machine learning approach to predict both NDI and continuous Bayley Scales of Infant and Toddler Development 3rd ed subscales at 2 year follow-up using: 1) the 5 variables used in the National Institute of Child Health and Human Development (NICHD) Extremely Preterm Birth Outcomes Tool, 2) 21 variables associated with outcomes in extremely preterm (EP) infants, and 3) a hypothesis-free approach using 133 potential variables available for infants in the PENUT database. Findings: The NICHD 5-variable model predicted 3–4% of the variance in the Bayley subscale scores, and predicted NDI with an area under the receiver operator curve (AUROC, 95% CI) of 0.62 (0.56–0.69). Accuracy increased to 12–20% of variance explained and an AUROC of 0.77 (0.72–0.83) when using the 21 pre-selected clinical variables. Hypothesis-free variable selection using BART resulted in models that explained 20–31% of Bayley subscale scores and AUROC of 0.87 (0.83–0.91) for severe NDI, with good calibration across the range of outcome predictions. However, even with the most accurate models, the average prediction error for the Bayley subscale predictions was around 14–15 points, leading to wide prediction intervals. Higher total transfusion volume was the most important predictor of severe NDI and lower Bayley scores across all subscales. Interpretation: While the machine learning BART approach meaningfully improved predictive accuracy above a widely used prediction tool (NICHD) as well as a model utilizing NDI-associated clinical characteristics, the average error remained approximately 1 standard deviation on either side of the true value. Although dichotomous NDI prediction using BART was more accurate than has been previously reported, and certain clinical variables such as transfusion exposure were meaningfully predictive of outcomes, our results emphasize the fact that the field is still not able to accurately predict the results of complex long-term assessments such as Bayley subscales in infants born EP even when using rich datasets and advanced analytic methods. This highlights the ongoing need for long-term follow-up of all EP infants. Funding: Supported by the National Institute of Neurological Disorders and Stroke U01NS077953 and U01NS077955.