Web-based newborn screening system for metabolic diseases: machine learning versus clinicians.

Area of Science:

• Biomedical Informatics
• Medical Technology
• Computational Biology

Background:

• Hospital Information Systems (HIS) integration of screening data is crucial but faces accuracy challenges.
• Disease characteristics and limited analytes impact the precision of traditional newborn screening methods.

Purpose of the Study:

• To enhance a neonatal screening system at National Taiwan University Hospital using a service-oriented architecture (SOA).
• To improve the accuracy of newborn screening for inborn errors of metabolism via machine learning and optimal feature selection.

Main Methods:

• Implemented a Newborn Screening Hospital Information System (NSHIS) utilizing HL7 standards and Web services.
• Employed machine learning (Support Vector Machine - SVM) with optimal feature selection for classifying phenylketonuria (PKU), hypermethioninemia, and 3-MCC deficiency.
• Analyzed 347,312 newborn blood samples, ranking 35 analytes and selecting top features for SVM models validated with 5-fold cross-validation.

Main Results:

• Optimal feature selection strategies identified key markers for PKU, hypermethioninemia, and 3-MCC deficiency.
• Machine learning approach substantially reduced false positive cases: PKU (21 to 2), hypermethioninemia (30 to 10), and 3-MCC deficiency (209 to 46).

Conclusions:

• The SOA Web service-based system efficiently accelerates newborn screening procedures.
• SVM methodology with optimal feature selection enhances classification accuracy for metabolic diseases.
• Adoption of these findings can dramatically decrease the number of suspected newborn screening cases.