Quantitative assessment of neonatal health using dried blood spot metabolite profiles and deep learning.

Summary

Using 13,536 preterm infants’ newborn dried blood spot metabolomics, the authors developed a deep-learning metabolic health index that stratifies risk for BPD, IVH, NEC, and ROP beyond gestational age and birthweight. The model outperformed other machine learning and clinical-variable models and was externally validated in 3,299 very preterm infants, reproducing biological risk subgroups.

Key Findings

• A deep-learning metabolic health index was derived from 13,536 newborn screening dried blood spot profiles linked to outcomes.
• The index stratified risk for BPD, IVH, NEC, and ROP independent of gestational age and birthweight.
• The model outperformed other machine-learning algorithms and clinical variable models.
• External validation in 3,299 very preterm infants reproduced common metabolic risk subgroups.

Clinical Implications

Early integration of the metabolic health index into neonatal workflows could refine risk stratification for preterm infants and enable targeted surveillance and interventions beyond gestational age/birthweight metrics.

Limitations

• Retrospective design relying on linked registry and screening data
• Generalizability may be constrained to regions using similar metabolite panels and screening workflows

Future Directions

Prospective implementation trials to assess workflow integration, cost-effectiveness, and impact on neonatal outcomes; mechanistic studies linking metabolite signatures to disease pathways.