Continual learning adapts adult chest X-ray AI for neonatal endotracheal tube assessment across 30 NICUs

Across 30 NICUs in 20 countries, privacy-preserving continual learning reduced neonatal chest X-ray endotracheal tube landmark error to 0.67 cm (vs 1.29 cm site-specific fine-tuning, 4.30 cm adult-only) at 29/30 sites and improved deep intubation detection at 27/30 sites.

• Design: international retrospective study with 30 NICUs across 20 countries/territories; each site contributed 50 neonatal chest radiographs without centralized data pooling.
• Intervention: continual learning adapted an adult deep learning endotracheal tube assessment model; comparators were hospital-specific fine-tuning and the adult-only model.
• Result: continual learning gave the lowest landmark localization error at 29/30 sites and improved deep intubation detection at 27/30 sites.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.