Implementing Explainable Machine Learning Models for Practical Prediction of Early Neonatal Hypoglycemia

Area of Science:

• Neonatal Medicine
• Computational Biology
• Medical Informatics

Background:

• Neonatal hypoglycemia is a prevalent metabolic disorder requiring early detection.
• Timely identification of at-risk neonates is crucial for optimizing neonatal care strategies.

Purpose of the Study:

• To develop and implement a machine learning model for predicting neonatal hypoglycemia risk.
• To create a predictive application integrated into a hospital information system for clinical decision support.

Main Methods:

• Retrospective analysis of 2687 neonates' electronic medical records (≥35 weeks gestational age).
• Evaluation of nine machine learning models using 12 clinical features.
• Selection of the best performing model based on Area Under the Receiver Operating Characteristic Curve (AUC).

Main Results:

• The Random Forest model demonstrated strong predictive performance (AUC=0.732) with good accuracy (0.658) and sensitivity (0.682).
• Key predictors identified include mode of delivery, gestational age, multiparity, respiratory distress, and low birth weight (<2500 gm).
• The best model was integrated into a web-based application within the hospital information system.

Conclusions:

• The developed machine learning model effectively assists clinicians in identifying neonates at risk of early hypoglycemia.
• Early identification enables timely interventions and treatment, potentially improving neonatal health outcomes.
• The predictive application serves as a valuable tool in neonatal intensive care units.