Machine learning models for reducing false positives in Fluorometric newborn screening of phenylketonuria:...

Area of Science:

• Biomedical informatics
• Neonatal screening
• Machine learning in healthcare

Background:

• Fluorometric newborn screening for phenylketonuria (PKU) is a standard diagnostic tool.
• High false positive rates in PKU screening lead to increased healthcare costs and parental anxiety.
• Machine learning (ML) offers potential to improve the accuracy and efficiency of newborn screening programs.

Purpose of the Study:

• To develop and validate machine learning models for reducing false positives in fluorometric PKU screening.
• To identify key clinical predictors for stratifying phenylketonuria risk in newborns.
• To create a practical tool for optimizing newborn screening workflows.

Main Methods:

• A retrospective analysis of 493,652 neonates' fluorometric PKU screening data was conducted.
• Predictors were identified using LASSO regression, and four ML models were trained and evaluated.
• Internal, test, and temporal validation cohorts, including propensity score matching, were used to assess model performance.

Main Results:

• Logistic regression demonstrated the best performance, with an AUC of 0.873 in internal validation and 0.955 in the test cohort.
• The model achieved 100% sensitivity in both test and temporal cohorts, effectively reducing false positives.
• Key predictors identified were phenylalanine concentration, gestational age, and birth weight.

Conclusions:

• A logistic regression model utilizing three routine clinical variables accurately stratifies PKU risk in screen-positive newborns.
• This ML-driven approach significantly reduces false positives without compromising screening sensitivity.
• The developed prediction tool provides a cost-effective solution for enhancing fluorometry-based newborn screening.