A newborn metabolic health index
Quantitative assessment of neonatal health using dried blood spot metabolite profiles and deep learning
Alan L. Chang, Jonathan D. Reiss, Martin Becker, Anthony Culos, Jonathan A. Mayo, Ivana Marić, Ramin Fallahzadeh, Neda H. Bidoki, Davide De Francesco, Thanaphong Phongpreecha, Camilo Espinosa, Maria Xenochristou, Natalie Stanley, Henry C. Lee, David K. Stevenson, Gary M. Shaw, Karl G. Sylvester, Nima Aghaeepour. Submitted for Review. 2021.
Motivation: Neonatal prematurity is associated with significant health risks and grave acquired conditions of prematurity that result in considerable morbidity and mortality. Historical gestational age and birthweight-based classifications of prematurity are limited in differentiating risk because similarly classified infants vary substantially in terms of health outcomes. Therefore, identifying risk metrics that can be used to stratify infants to appropriate care pathways is an urgent clinical need.
Results: A comprehensive metabolic newborn risk assessment was performed by linking quantitative metabolite results from over 13,000 newborn screening (NBS) blood spot tests to clinical outcomes of prematurity. Using deep learning and subgroup discovery approaches, an NBS-based metabolic health index was developed that identifies the likelihood of suffering from four major adverse neonatal outcomes in extremely preterm infant populations. Thus, NBS and artificial intelligence approaches can potentially power a new, more granular biological definition of prematurity to help the most vulnerable neonatal patients.
Data Availability: Pre-existing data access policies outlined by the California State Biobank, the California Office of Statewide Health Planning and Development (OSHPD), and the California Perinatal Quality Care Collaborative (CPQCC) govern data access requests. Requests will be reviewed by the steering committees from each organization prior to providing access.
Code Availability: The model developed as part of this work (applicable to new datasets) as well as the model training scripts will be publicly available through Github upon publication of the article. Anonymous access for the peer review process is available here.