Adapting an adult chest-radiograph model to neonatal intensive care across 30 units
A model trained to assess endotracheal tube position on adult chest radiographs was retrained in sequence at 30 neonatal intensive care units in 20 countries and territories, 50 radiographs per site and with no pooling of patient data, reaching a mean landmark localisation error of 0.67 cm against 1.29 cm for hospital-specific fine-tuning and 4.30 cm for the unadapted adult model.
- Model
- Adult endotracheal-tube assessment deep learning model adapted by continual learning; the base model is not named in the abstract
- Field
- Medicine
- Date
- 2026-10-13
- Human collaborators
- Vishwanatha M Rao, Pranav Rajpurkar
What was found
Fewer than one in five commercially available radiology AI models are cleared for paediatric use, models trained on adults generalise poorly to neonates, and neonatal images are scarce in public datasets, so the gap is usually closed by local data curation that most units cannot do. This international retrospective study evaluated continual learning as a privacy-preserving way to adapt an adult endotracheal-tube assessment model to neonatal care: each of 30 neonatal intensive care units across 20 countries and territories contributed 50 radiographs, and the model was retrained sequentially from site to site with no centralised pooling. Continual learning reached a mean anatomical landmark localisation error of 0.67 cm, against 1.29 cm for hospital-specific fine-tuning and 4.30 cm for the adult-only model, and was better at 29 of the 30 sites. Detection of deep intubations improved at 27 of 30 sites. The authors present this as evidence that collaborative learning without data sharing can address a paediatric deployment gap, not as a clinical outcome result.
Novelty check
Searched Europe PMC on 2026-10-22 (TITLE:"continual learning" AND TITLE:neonatal, which returns this paper alone, and "endotracheal tube" AND "continual learning", which returns this paper plus a governance review and no other multicentre neonatal adaptation study). The nearest prior work the searches surfaced is federated continual learning for chest-radiograph classification in Scientific Reports (doi 10.1038/s41598-026-55211-7), a classification task rather than neonatal tube placement. No prior registry entry covers paediatric radiology AI or continual learning, and none of the registry's existing imaging entries concerns neonates.
Caveats and known objections
Retrospective throughout, with no prospective deployment and no clinical outcome: the endpoints are localisation error in centimetres and deep-intubation detection, both surrogates for correct tube placement rather than measures of harm avoided. Fifty radiographs per site is a small adaptation set, and the 30 units were not sampled to be representative of neonatal intensive care anywhere. The base adult model is not named in the abstract. A Chinese-language reprint of the paper identifies it as CarinaNet, a RetinaNet detector described in Critical Care in 2023 (doi 10.1186/s13054-023-04320-0), which would make it an academic rather than a commercial system, but the paper itself was not readable for this entry, so that identification is unconfirmed, as are the per-site figures the same reprint gives, several of which disagree with the published abstract. Sponsor and vendor roles, the funding statement and the authors' own limitations section are therefore unchecked here. Autonomy is ai-assisted: the model localises anatomical landmarks for a clinician assessing tube position, and no care decision was measured. The paper is not open access and the abstract does not say whether the site data or the adapted weights will be released, so a reader outside the consortium cannot obtain either.
Nobody outside the lab has checked this yet.
Reading the primary source closely enough to say whether it supports the claim counts as a check, and you are credited on the entry.
Or on GitHub: submit a check challenge the grade send a correction or send a pull request
Flag this for triage
Signals order the review queue and nothing else. They are never published, and they never move a grade: that takes a citation.
Entry history (1 event)
- AddedEntered the registry graded Peer reviewed and AI-assisted.
Entries are never deleted. A grade that does not hold up is downgraded on the record, with the reason beside it.
Graded peer reviewed for verification and ai-assisted for autonomy. What these mean.
Cite this entry
whataifound.org. (2026). Adapting an adult chest-radiograph model to neonatal intensive care across 30 units. whataifound.org: A Registry of AI Scientific and Mathematical Discoveries. https://whataifound.org/finding/2026-10-13-neonatal-radiograph-continual-learningBibTeX
@misc{whataifound-internationalnicucontinuallearningconsortium-2026-learning,
title = {Adapting an adult chest-radiograph model to neonatal intensive care across 30 units},
author = {{whataifound.org}},
year = {2026},
howpublished = {whataifound.org: A Registry of AI Scientific and Mathematical Discoveries},
note = {Result by International NICU continual-learning consortium. Verification: Peer reviewed. Autonomy: AI-assisted.},
url = {https://whataifound.org/finding/2026-10-13-neonatal-radiograph-continual-learning}
}