Ensemble learning for the early prediction of neonatal jaundice with genetic features

Abstract

Background: Neonatal jaundice may cause severe neurological damage if poorly evaluated and diagnosed when high bilirubin occurs. The study explored how to effectively integrate high-dimensional genetic features into predicting neonatal jaundice.

Methods: This study recruited 984 neonates from the Suzhou Municipal Central Hospital in China, and applied an ensemble learning approach to enhance the prediction of high-dimensional genetic features and clinical risk factors (CRF) for physiological neonatal jaundice of full-term newborns within 1-week after birth. Further, sigmoid recalibration was applied for validating the reliability of our methods.

Results: The maximum accuracy of prediction reached 79.5% Area Under Curve (AUC) by CRF and could be marginally improved by 3.5% by including genetic variant (GV). Feature importance illustrated that 36 GVs contributed 55.5% in predicting neonatal jaundice in terms of gain from splits. Further analysis revealed that the main contribution of GV was to reduce the false-positive rate, i.e., to increase the specificity in the prediction.

Conclusions: Our study shed light on the theoretical and practical value of GV in the prediction of neonatal jaundice.

Keywords: Genetic variants; Hyperbilirubinemia; Machine learning; Transcutaneous bilirubin.

Fig. 1

The methodological workflow of our study

Fig. 2

The architecture of Gradient Boosting Decision Tree

Fig. 3

Relative feature importance from ensemble nethod in predicting neonatal jaundice under CN220 guideline

Fig. 4

Calibration curves on external validation sets

Fig. 5

ROC curve of neonatal jaundice prediction with CRF and GV by ensemble learning

Fig. 6

Comparison of ROC curve of neonatal jaundice prediction after introducing genetic variants (GV36)