Adult intubation AI retooled for newborns: 30 NICUs pass the model along, cutting landmark error from 1.29 cm to 0.67 cm

Endotracheal tube position assessment on neonatal chest radiographs can now be handled by a model that starts from adult data, without training a new model at every hospital. Adult models previously lost substantial performance on neonates, while adapting them with only local data was limited by the small number of annotated images each hospital held. In a retrospective cohort from 30 neonatal intensive care units in 20 countries and territories, continual learning averaged 0.67 cm combined tube tip and carina localization error, against 1.29 cm for hospital-specific fine-tuning and 4.30 cm for direct prediction; localization improved at 29 of 30 sites and reached statistical significance at 27.

The model moves from hospital to hospital, and each site retrains it on 50 chest radiographs locally before handing the updated weights to the next site. This route avoids pooling patient images centrally and avoids the synchronized distributed training infrastructure that federated learning requires. The study covered 2,965 neonatal chest radiographs from 2,960 patients, with each hospital contributing exactly 50 training images; results were averaged over 10 random hospital orderings, and the trained model weights are publicly available.

Sequential training did not make the model forget earlier hospitals; performance kept improving as more sites joined. The usual risk in continual learning is that later updates overwrite earlier knowledge and degrade performance at sites seen first. Across 100 random orderings, comparing each hospital's intermediate model right after its own training with the final model after the full sequence showed a mean further reduction of 0.48 cm in combined localization error at sequence position 0, with the gain shrinking for later positions (Spearman ρ = −0.85, P < 0.001).

This is a retrospective proof of concept spanning 30 neonatal intensive care units in 20 countries and territories, with patients aged 0 to 6 months and predominantly preterm. All experiments ran centrally to simulate decentralized execution, and each hospital contributed about 100 radiographs, 50 of them for training. Model weights are public, but the radiographs are not, because the data use agreements prohibit redistribution. The authors propose next integrating the model into neonatal intensive care workflows at a subset of participating sites and prospectively measuring time to confirmation for abnormal tube placements, clinician agreement with model-flagged cases and downstream re-intubation rates.

Hospitals were asked to supply class-balanced datasets, so performance could shift with the true prevalence of abnormal tube placements. The pixel-to-centimetre conversion relies on estimated tube width, which probably contributes to imperfect correlation between tube-to-carina distance and clinical labels, and the correspondence is markedly weaker for high placements than for low ones. One hospital with low-resolution radiographs was excluded from the main analysis; including it raised combined localization error from 0.67 cm to 0.75 cm, suggesting data quality heterogeneity can constrain the benefit. The authors also note that continual model updates may complicate regulatory oversight, and that extremely premature neonates have a narrower margin for acceptable tube position, so prospective validation is needed to determine whether the observed accuracy is clinically sufficient.