NeoGx: Machine-Recommended Rapid Genome Sequencing for Neonates

Machine learning algorithm using electronic health record data, including phenotypes derived from clinical text, to predict early which neonates will require genetic evaluation within the first 18 months of life.

Genetic disease is common in Level IV neonatal intensive care units, yet clinicians struggle to identify which newborns would benefit from genetic evaluation. NeoGx is a machine learning algorithm developed on electronic health record data from 14,272 Level IV NICU patients — structured data and phenotypes derived from clinical text — temporally split into development (n = 11,201), calibration (n = 1,080) and validation (n = 1,991) cohorts, designed to predict early which neonates will require genetic evaluation within the first 18 months of life. Using predictions accumulated over four NICU weeks, NeoGx reaches a ROC AUC of 0.849 and a precision-recall AUC of 0.771 in the independent validation cohort. NeoGx-guided referral would reduce the mean time to first genetic evaluation from 44 to 29 days, an average gain of 15.2 days; paired with rapid genome sequencing as the first-line test, the share of genetic cases reaching a definitive testing endpoint within 14 days would rise from 9.5% to 68.6%.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

The idea is sound and rarely addressed: the limiting factor for rapid genome sequencing in neonatal intensive care is not the technology but recognition of the indication, and a ROC AUC of 0.849 in a temporally independent validation cohort is a credible starting point. The main reservation concerns the most striking figures: the rise from 9.5% to 68.6% of cases reaching a definitive endpoint within 14 days is a counterfactual estimate of what algorithm-guided referral would have produced, not the result of a prospective implementation. What remains to be shown is robustness in a real clinical flow, acceptability among neonatologists and portability to other electronic health record systems, since models fed by clinical text depend heavily on local documentation habits.

Analysis by Dr Thibaut Benquey