Predictive models for neonatal follow-up serum bilirubin: Model development and validation

Abstract

Background: Hyperbilirubinemia affects many newborn infants and, if not treated appropriately, can lead to irreversible brain injury. Objective: This study aims to develop predictive models of follow-up total serum bilirubin measurement and to compare their accuracy with that of clinician predictions. Methods: Subjects were patients born between June 2015 and June 2019 at 4 hospitals in Massachusetts. The prediction target was a follow-up total serum bilirubin measurement obtained <72 hours after a previous measurement. Birth before versus after February 2019 was used to generate a training set (27,428 target measurements) and a held-out test set (3320 measurements), respectively. Multiple supervised learning models were trained. To further assess model performance, predictions on the held-out test set were also compared with corresponding predictions from clinicians. Results: The best predictive accuracy on the held-out test set was obtained with the multilayer perceptron (ie, neural network, mean absolute error [MAE] 1.05 mg/dL) and Xgboost (MAE 1.04 mg/dL) models. A limited number of predictors were sufficient for constructing models with the best performance and avoiding overfitting: current bilirubin measurement, last rate of rise, proportion of time under phototherapy, time to next measurement, gestational age at birth, current age, and fractional weight change from birth. Clinicians made a total of 210 prospective predictions. The neural network model accuracy on this subset of predictions had an MAE of 1.06 mg/dL compared with clinician predictions with an MAE of 1.38 mg/dL (P