Primary risk stratification for neonatal jaundice among term neonates using machine learning algorithm

Abstract

Background: Neonatal jaundice occurs in approximately 60% of term newborns. Although risk factors for neonatal jaundice have been studied, all the suggested strategies are based on various newborn tests for bilirubin levels. We aim to stratify neonates into risk groups for clinically significant neonatal jaundice using a combined data analysis approach, without serum bilirubin evaluation.

Study design: Term (gestational week 37-42) neonates born in a single medical center, 2005-2018 were identified. Anonymized data were analyzed using machine learning. Thresholds for stratification into risk groups were established. Associations were evaluated statistically using neonates with and without clinically significant neonatal jaundice from the study population.

Results: A total of 147,667 consecutive term live neonates were included. The machine learning diagnostic ability to evaluate the risk for neonatal jaundice was 0.748; 95% CI 0.743-0.754 (AUC). The most important factors were (in order of importance) maternal blood type, maternal age, gestational age at delivery, estimated birth weight, parity, CBC at admission, and maternal blood pressure at admission. Neonates were then stratified by risk: 61% (n = 90,140) were classed as low-risk, 39% (n = 57,527) as higher-risk. Prevalence of jaundice was 4.14% in the full cohort, and 1.47% and 8.29% in the low- and high-risk cohorts, respectively; OR 6.06 (CI: 5.7-6.45) for neonatal jaundice in high-risk group.

Conclusion: A population tailored "first step" screening policy using machine learning model presents potential of neonatal jaundice risk stratification for term neonates. Future development and validation of this computational model are warranted.

Keywords: Machine learning; Neonatal jaundice; Obstetrics; Personalized medicine; Prediction.