Integration of Machine Learning Techniques as Auxiliary Diagnosis of Inherited Metabolic Disorders: Promising Experience with Newborn Screening Data

Abstract

Tandem mass spectrometry is an advanced biochemical analysis method and has been widely used in screening of inherited metabolic disorders (IMDs). Obtained examination results are filtered by cutoff values and then interpreted based on doctor’s knowledge to get diagnoses. However, cutoff-based approaches have difficulties with the correlations of multiple metabolites. Doctor’s experiences affect the diagnostic decision-making as well. The rapidly increasing availability of newborn screening data (1.5M cases in this study) enables the application of machine learning (ML) techniques to provide more accurate diagnoses of IMDs compared to simple cutoff values. We investigated two tasks in this study, i.e. complicated patterns between metabolites and better auxiliary diagnostic means. Experimental results show that novel metabolic patterns found in the study are effective and meaningful. Integrating ML techniques with these patterns improved predictive performance compared to existing diagnostic methods, suggesting ML techniques are becoming valuable as auxiliary diagnostic tools.