Prediction of Neurodevelopmental Outcomes in Very Preterm Infants: Comparing Machine Learning Methods to Logistic Regression.

PURPOSE

Is machine learning (ML) superior to the traditionally used logistic regression (LR) in prediction of neurodevelopmental outcomes in preterm infants?

OBJECTIVES

To develop and internally validate a ML model to predict neurodevelopmental impairment (NDI) in very preterm infants (<31 weeks) at 36 months corrected age, using clinical predictors.

METHODS

A retrospective cohort of very preterm infants (2330 weeks) born between January 2004 and December 2016 in Nova Scotia, Canada. Survivors with neurodevelopmental assessment at 36 months corrected age were included. The study sample was randomly split (80:20) into a development and testing datasets. We compared four methods: LR, elastic net (EN), random forest ensemble (RF) and gradient boosting (XGB), in relation to discrimination (AUC), calibration, and diagnostic properties.

RESULTS

Of 811 eligible infants, 663 were included (mean gestational age 28 weeks, mean birth weight 1137 g and 52% male). Of those, 195 (29%) developed NDI and 468 (71%) did not. On internal validation using the testing dataset, all four models provided good discrimination of NDI with comparable AUC. RF was superior to the other three methods with a higher AUC (0.79 vs. 0.74, 0.74, and 0.73 for XGB, EN and LR, respectively), but all models have overlapped CIs.

CONCLUSIONS

In this population-based cohort of very preterm infants, RF was superior to conventional LR in prediction of NDI at 3 years corrected age. Accurate prediction of preterm infants at risk of NDI enables early referrals for intervention programs and resources allocation toward those who are most likely to benefit.