Validating the early phototherapy prediction tool across cohorts

Area of Science:

• Neonatal Medicine
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Neonatal hyperbilirubinemia is a common global condition.
• Early detection and treatment are crucial for preventing adverse outcomes.
• A machine learning tool (EPPT) was previously developed for early phototherapy prediction.

Purpose of the Study:

• To evaluate the applicability and performance of the EPPT on a new, distinct patient cohort.
• To validate the predictive accuracy of the EPPT in a different population.
• To compare the original EPPT model with a re-trained version on new data.

Main Methods:

• Retrospective analysis of prospectively recorded neonatal data from 1,109 infants.
• Application of the original EPPT (logistic regression and random forest ensemble).
• Re-training of the EPPT model on the new cohort data and comparative analysis of predictive performance.

Main Results:

• The original EPPT achieved 84.6% AUROC for phototherapy prediction in the new cohort.
• Re-training the model resulted in an improved AUROC of 88.8% via cross-validation.
• Gestational age, birth weight, bilirubin-to-weight ratio, hours since birth, and bilirubin value were key predictors.

Conclusions:

• The EPPT demonstrates robust predictability of treatment for neonatal hyperbilirubinemia across diverse patient cohorts.
• The tool requires only a single total serum bilirubin measurement and four clinical parameters.
• Further prospective studies are recommended to develop a clinical decision support system.