Predictive Models for Neonatal Follow-Up Serum Bilirubin: Model Development and Validation

Area of Science:

• Neonatal Medicine
• Artificial Intelligence in Healthcare
• Predictive Analytics

Background:

• Neonatal hyperbilirubinemia is a common condition in newborns.
• Untreated, it can lead to severe, irreversible brain injury.

Purpose of the Study:

• To develop and compare the accuracy of predictive models for follow-up total serum bilirubin (TSB) measurements.
• To evaluate these models against clinician predictions.

Main Methods:

• Supervised learning models were trained on TSB measurements from 4 Massachusetts hospitals (June 2015-June 2019).
• A dataset of 27,428 measurements was used for training and 3320 for testing.
• Model performance was compared to prospective clinician predictions.

Main Results:

• Neural network and Xgboost models achieved the highest predictive accuracy (MAE 1.05 and 1.04 mg/dL, respectively).
• Key predictors included current bilirubin, rate of rise, phototherapy use, and gestational age.
• AI models significantly outperformed clinician predictions (MAE 1.06 vs. 1.38 mg/dL, P<.0001).

Conclusions:

• Developed predictive models for neonatal TSB measurements demonstrate superior accuracy compared to clinicians.
• These models are the first to predict specific bilirubin values, include diverse patient populations, and account for phototherapy effects.