Machine learning models for reducing false positives in Fluorometric newborn screening of phenylketonuria: Development, validation, and clinical application.

BACKGROUND

Fluorometric newborn screening for phenylketonuria (PKU) is widely used but plagued by high false positives (FPs), leading to unnecessary follow-ups and parental anxiety. Machine learning (ML) models may optimize screening efficiency by reducing FPs while preserving sensitivity.

METHODS

A retrospective study used fluorometric PKU screening data from 493,652 neonates (2022-2024) at Shenzhen Maternity and Child Healthcare Hospital. The cohort included 70,342 concurrent tandem mass spectrometry (MS/MS) tests, with 711 screen-positive cases selected for internal validation. A 1:2 propensity score matching (PSM) cohort (n = 120, 40 confirmed PKU cases and 80 false-positive cases) was constructed to validate model robustness. An temporal cohort comprised 539 screen-positive cases from 154,721 neonates screened in 2021. Key predictors were selected via least absolute shrinkage and selection operator (LASSO) regression. Four ML models (logistic regression, extreme gradient boosting (XGBoost), support vector machine (SVM), random forest) were trained and evaluated based on sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the curve (AUC).

RESULTS

LASSO regression identified three predictors: fluorometric phenylalanine (Phe) concentration, gestational age, and birth weight. Logistic regression performed best, achieving an AUC of 0.873 (95% CI: 0.780-0.958) in internal validation and 0.955 in the test cohort. Temporal validation confirmed strong performance, with an AUC of 0.927 and 100% sensitivity in both test and temporal cohorts. In the PSM test subset, the model yielded an AUROC of 0.813 and a PPV of 0.538. A freely accessible web-based prediction tool was developed and deployed.

CONCLUSION

Using only three routine clinical variables, our logistic regression model effectively stratified PKU risk in screen-positive newborns, significantly reducing FPs without sacrificing sensitivity. This offers a cost-effective tool for optimizing fluorometry-based newborn screening workflows.