Published on October 13, 2026 in Nature Biomedical Engineering, a international work shows that a model designed to analyze the Placement of an endotracheal tube in adults can be adapted to newborns by continuous learning, without grouping all the X-rays into a single central base. Performance has improved significantly in 30 neonatal intensive care units, but the study remains retrospective and does not yet demonstrate a clinical benefit in real-world settings.[1]
What Just Happened
A team of researchers and clinicians from several continents evaluated the adaptability of a chest X-ray model originally developed for adults. Its purpose is to help locate an endotracheal tube and identify cases of intubation that are too deep, a situation that can expose newborns to respiratory complications.[1][2]
The study is based on X-rays from 30 neonatal intensive care units located in 20 countries and territories. Each facility provided 50 images, for a total of 1,500 X-rays. The data were not collected in a single repository. The model was retrained successively using data from each site.[1][2]
The context justifies this approach. An analysis published in JAMA Pediatrics identified 876 medical devices using AI or machine learning that have been approved by the U.S. Food and Drug Administration. Only 149—about 17%—were explicitly approved for use in a pediatric population. Of these 149 devices, only 28 mentioned the inclusion of pediatric patients in their validation data.[3]
What’s Really New
Using an adult model as a starting point is nothing new. The challenge lies in adapting it to a population whose anatomy, body dimensions, medical conditions, and acquisition conditions differ significantly. A neonatal X-ray is not simply a scaled-down version of an adult X-ray.
The study’s contribution lies in the combination of two choices. The first is to build upon an existing model rather than develop a neonatal system from scratch. The second is based on continuous learning: the model learns incrementally from small datasets collected at several hospitals.
This method addresses a practical challenge. Neonatal units rarely have thousands of annotated images, and the transfer of health data between countries raises legal, technical, and ethical challenges. Continuous learning therefore seeks to pool the experience of these institutions without directly sharing their images.
How Does Continuous Learning Work
In a traditional centralized training process, data from multiple hospitals is copied into a shared environment. The model is then trained on this dataset. Here, the process is sequential: the model is trained on the 50 X-rays from one facility, then continues learning with those from the next facility.
- The adult model provides an initial representation of a chest X-ray and an endotracheal tube.
- Each neonatal unit contributes a small local set of annotated images.
- The model parameters are updated sequentially, while the X-rays remain at their original facility.
- The final system is evaluated at each of the 30 sites to assess its ability to adapt to different contexts.
This structure reduces the need to centralize images, but it is not enough to guarantee absolute confidentiality. A model’s parameters and updates can themselves disclose information if they are not properly protected. Security therefore also depends on the transmission protocol, access controls, update auditing, and the mechanisms used to defend against attacks targeting the models.
Why an adult model does not automatically become universal
Generalization refers to a model’s ability to remain reliable when applied to data different from that used during its training. In medicine, this ability may deteriorate when patients’ ages, the prevalence of diseases, or a facility’s equipment or practices change.
An analysis published in 2026 in *Nature Health* examined 203 publicly available medical imaging datasets. Patient age was missing in one-third of the datasets examined, and children accounted for less than 2% of patients when this information was available. Furthermore, 95.6% of the pediatric patients identified were from the United States. This concentration limits the representation of the anatomy, practices, and equipment found in other regions.[5]
Newborns present a particularly challenging case. Their small size alters anatomical landmarks. Premature infants may present specific clinical conditions, while X-ray equipment and protocols vary from one facility to another. Therefore, high average performance in adults does not provide sufficient information about reliability in these patients.
What the results actually allow us to conclude
The model adapted using continuous learning achieved an average error in the localization of anatomical landmarks of 0.67 cm. This error was 1.29 cm when adjustments were made separately at each hospital and 4.30 cm when the adult model was used without adaptation. The continuous approach reduced the error at 29 out of 30 sites and improved the detection of deep intubations at 27 out of 30 sites.[1][2]
These figures represent a technical improvement over the study data. They do not yet demonstrate that the system reduces intubation errors, speeds up care, or improves the health status of newborns. The study is retrospective and does not measure the model’s effect on clinicians’ real-time decisions.
Liability Issues
The first requirement concerns local validation. A model may work well in most facilities but still be inadequate at a specific site. Before any deployment, each hospital should verify performance on its own equipment, imaging protocols, and patient population.
The second challenge relates to the model’s evolution. A system that is trained successively can learn new contexts while losing some of its previous performance—a phenomenon often referred to as catastrophic forgetting. Robust governance must therefore monitor performance after each update, maintain auditable versions, and provide for a rollback to a previous version in the event of a performance decline.
Data protection should not be limited to the location of the data. The fact that X-rays do not leave the hospital reduces certain risks, but the circulation of the model creates other responsibilities: identifying those authorized to modify it, securing data exchanges, ensuring the traceability of contributions, and verifying that no information has been leaked.
Finally, the model applies in situations where patients are unable to give consent themselves. Legal representatives must be provided with understandable information, while healthcare facilities must justify the use of the data and ensure that the radiologist, neonatologist, and healthcare team remain accountable. In the European Union, AI integrated into a medical device may fall under the high-risk system regime when the conditions of the AI Act are met.[6]
What to Watch for Now
The next step will be prospective. We will need to evaluate how the system performs during actual patient care, using images acquired in real time, decisions made under time pressure, and a full range of clinical situations. The relevant criteria will then extend beyond localization errors to include the rate of relevant alerts, correction time, complications avoided, the burden on staff, and incidents related to the model.
We will also need to examine the stability of continuous learning. The order in which institutions participate, the quality of the annotations, and differences between devices can all influence the final model. A collaborative method is equitable only if it improves performance in the least-represented contexts without compromising the performance already achieved elsewhere.
This study therefore provides a partial but solid answer to the question posed. Yes, an adult model can be adapted for use in neonatology without centralizing all images—at least for a specific radiographic task and in a retrospective evaluation. The clinical impact of this strategy will now depend on its prospective validation, its safety, and the ability of institutions to manage a model that continues to evolve.
For more information
To learn more about the challenges of widespread adoption, clinical oversight and medical data governance , check out these analyses published on the aivancity blog.
Sources
[1] Rao, V. M., Chen, E., Saenz, A., et al. (2026). Repurposing adult AI radiograph models for neonatal care through continual learning: an international multicenter study. Nature Biomedical Engineering. Published October 13, 2026. View the study
[2] National Library of Medicine. PubMed, record PMID 42869977. Accessed October 14, 2026. View record
[3] Brewster, R. C. L., Nagy, M., Wunnava, S., and Bourgeois, F. T. (2025). US FDA Approval of Pediatric Artificial Intelligence and Machine Learning-Enabled Medical Devices. JAMA Pediatrics, 179(2), 212–214. View publication
[4] Shin, H. J., Son, N. H., Kim, M. J., & Kim, E. K. (2022). Diagnostic performance of artificial intelligence approved for adults for the interpretation of pediatric chest radiographs. Scientific Reports, 12, 10215. View the study
[5] Hua, S. B. Z., Heller, N., He, P., et al. (2026). Underrepresentation of children in public medical imaging datasets. Nature Health, 1, 888–897. View the study
[6] European Commission. (2026). Artificial Intelligence in Health. Accessed October 14, 2026. View source
