Can an adult X-ray AI learn neonatal care without moving patient data?
Across 30 neonatal intensive-care units, sequential local retraining cut average anatomical landmark error to 0.67 cm, compared with 1.29 cm for one-hospital fine-tuning and 4.30 cm for the unadapted adult model. The retrospective study did not deploy the system or measure harm, treatment changes or infant outcomes.
The short version
3 key points
- 1.Thirty NICUs in 20 countries and territories each contributed about 100 neonatal chest radiographs; 50 per hospital were used for local training and 43 to 53 for testing.
- 2.Continual learning produced a combined landmark-localisation error of 0.67 cm, versus 1.29 cm for hospital-specific fine-tuning and 4.30 cm when the adult model was used without neonatal adaptation; it improved performance at 29 of 30 sites.
- 3.The images were retrospectively selected and deliberately balanced toward correct and incorrect tube positions. No model ran in a live NICU, no treatment decision was automated and no infant outcome was measured.
The answer: international local retraining worked technically, but clinical safety is unproven
A model first trained to locate an endotracheal tube tip and the carina on adult intensive-care radiographs performed poorly when applied directly to newborn images. The international team then moved the model weights—not the patient images—through a sequence of neonatal units. Each hospital retrained the model on 50 locally held radiographs before passing updated parameters onward. Across 30 sites, the final continual-learning approach reduced the combined average landmark-localisation error to 0.67 centimetres. Hospital-specific fine-tuning averaged 1.29 cm and the unadapted adult model 4.30 cm. Continual learning improved localisation at 29 of 30 sites and improved deep-intubation detection at 27 sites.
That is a meaningful technical result because a misplaced breathing tube in a newborn can require urgent review. It is not evidence that the system is ready to advise clinicians. The investigators analysed historical images with expert annotations; they did not place the model inside a neonatal workflow, measure whether it altered radiologists' or clinicians' decisions, record false reassurance during real emergencies, or test infant outcomes. The study shows that small, distributed datasets can adapt one specific adult imaging model. It does not establish that privacy is guaranteed, that every paediatric imaging task will transfer, or that an automated placement judgement is safe enough to use without specialist oversight.[1][2]
What the 30 hospitals contributed
The retrospective collaboration covered 30 neonatal intensive-care units in 20 countries and territories. Each site aimed to provide roughly 100 single bedside chest, chest-abdomen or anteroposterior radiographs from patients aged from birth to six months, including premature infants. The main collection year was 2021; when a centre did not have enough qualifying cases, earlier years could be included. All clinical indications were eligible. The target was pragmatic: broad geographic participation with a dataset small enough for institutions to curate locally.
Each hospital's training set contained exactly 50 images. The remaining 43 to 53 viable images formed its test set, giving roughly 3,000 radiographs across the consortium rather than a large population-level cohort. Sites aimed for about half correctly positioned and half incorrectly positioned tubes, using original radiology reports or collaborating physicians' review. The researchers also stratified the split across high, normal and low tube-placement categories. That balancing helps a model encounter uncommon mistakes, but it differs from the prevalence clinicians see in routine practice. Predictive values and workload in a naturally imbalanced NICU stream may therefore differ from this benchmark.[1]
How continual learning was compared with local and adult-only models
The starting system was an adult endotracheal-tube model. Direct prediction applied it to neonatal images unchanged. The second comparator fine-tuned a separate copy on each hospital's 50 training radiographs. The continual-learning system instead accumulated adaptation sequentially: hospital one updated the weights, hospital two began from that update, and so on. For every test site, the model was constructed from the other hospitals rather than that site's own training data, and the researchers averaged results across ten random hospital orderings to reduce the risk that one convenient sequence drove the finding.
The evaluation asked the model to localise the tube tip and the carina and then derive the distance used to classify placement. The released repository includes the preprocessing, sequential-training and leave-one-hospital-out scripts, while the Zenodo record preserves the software package. This is stronger transparency than a paper with no inspectable implementation. It still does not reproduce the result independently: the patient images remain local, and external teams cannot reconstruct the full international experiment from code alone. Updated model parameters can also leak information in some settings, so 'images never left the hospital' should not be read as proof that every privacy attack has been excluded.[1][2]
Why the improvement matters—and where averages can mislead
The headline comparison combines errors for two landmarks. A smaller error can produce a more reliable tip-to-carina distance, supporting earlier identification of a tube inserted too deeply. Continual learning outperformed local fine-tuning at nearly every site despite never training on the target hospital's own 50 images in the held-out evaluation. That suggests accumulated variation across countries and equipment may be more useful than a very small local sample, and that collaboration can help hospitals whose paediatric data are scarce.
The denominator is nevertheless modest at each hospital, and mean error does not describe every dangerous miss. A system can average 0.67 cm while occasionally being wrong by much more. Selection toward a balanced set also leaves calibration under real prevalence unresolved. Performance by gestational age, birth weight, device type, equipment, image quality and clinical subgroup needs enough cases and confidence intervals to expose uneven failure. The study's international breadth is valuable, but geographic coverage is not the same as a prospective test of local workflows, staffing, image acquisition and patient mix.[1]
What this could change for infants, clinicians and hospitals
For an infant, the possible benefit is faster recognition of a tube that sits too high or too low, without waiting for a hospital to assemble thousands of labelled neonatal images. For clinicians, a dependable flag could prioritise a radiograph for review rather than replace interpretation. The safe role at this stage is decision support with visible measurements, access to the original image and an easy route to override the model. A restored or annotated image should never hide uncertainty or make a misplaced landmark look authoritative.
For hospitals, the study offers a collaboration design: curate a small local dataset, train inside institutional infrastructure and transfer parameters rather than raw images. Deployment would still require privacy threat modelling, cybersecurity controls, version tracking, quality checks after every update and a clear party responsible for the final model. The work was supported by the Biswas Family Foundation's Transformative Computational Biology Grant in collaboration with the Milken Institute; one author also reported support from Junior Orsi and Barco. One author disclosed honoraria from several medical companies and shares in two firms; the other authors declared no competing interests.[1][2][3]
What would change the assessment
The next step should be a locked prospective shadow study: run the model on consecutive NICU radiographs without allowing it to influence care, then compare its alerts with independent expert review and final clinical decisions. Researchers should report sensitivity, specificity, calibration and predictive values at natural tube-placement prevalence, along with the largest localisation errors and subgroup results. The evaluation should include units absent from the original consortium and deliberately stress different scanners, exposure settings, infant sizes and tubes obscured by other devices.
Only after that should an interventional study ask whether the system shortens time to correction without increasing unnecessary repositioning, repeat imaging or alarm fatigue. Privacy testing should examine whether shared parameters expose training information and whether secure aggregation or other protections are needed. Evidence of fewer harmful delays, stable performance after updates and an auditable response to rare failures would move the assessment from promising technical adaptation toward clinical usefulness. Until then, the result supports research and controlled evaluation, not autonomous neonatal care.[1]
What this means for people
- Infants could benefit if a reliable flag speeds review of a misplaced breathing tube, but no infant outcome or live treatment decision was measured.
- Clinicians need the original radiograph, visible measurements and an override route; the model should prioritise review rather than replace it.
- Hospitals can collaborate without centralising images, while still needing privacy testing, cybersecurity controls and accountability for model updates.
Global context
The study's 30 NICUs span 20 countries and territories, an unusually broad test for neonatal imaging AI. Each centre contributed only about 100 deliberately balanced radiographs, however, so the work does not estimate real-world event prevalence or prove readiness across all equipment, staffing models and newborn populations.
What the evidence does not yet show
- The retrospective benchmark did not deploy the system, change care or measure safety and patient outcomes.
- Each hospital contributed only 43 to 53 test images, limiting estimates of rare and severe failure.
- The deliberately balanced placement categories do not reflect routine prevalence and leave real-world calibration unresolved.
- No independent team can reproduce the full experiment without access to the locally held images.
- Keeping images local reduces data movement but does not by itself prove resistance to information leakage from shared parameters.
What to watch next
- Prospective shadow-mode evaluation on consecutive NICU radiographs
- Worst-case errors and subgroup results by infant size, gestational age, equipment and site
- Privacy attacks and secure parameter-sharing safeguards
- Time to tube correction, repeat imaging, unnecessary intervention and infant outcomes
Living evidence record
Impact record IAI-1S427RZ
Evidence stage
Studied
Confidence
Corroborated
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
14 October 2026
Source trail
3 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 14 October 2026
Research question, topics and regions
Whether an adult endotracheal-tube assessment model can be adapted to neonatal chest radiographs by sequential local retraining across hospitals without pooling patient images
Topics: Neonatal care · Medical imaging · Continual learning · Data privacy · Endotracheal tubes · Clinical AI
Regions: Global · North America · Europe · Asia & Middle East · Africa · South America · Oceania
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Health & Life Sciences
Is neonatal respiratory AI ready for routine care?
A peer-reviewed scoping review mapped 35 studies using AI and digital tools to predict, image or monitor newborn breathing problems. The evidence spans promising prototypes, but heterogeneity, small samples, limited external validation and few clinical-impact tests keep routine use unproven.
9 min · 2 sources
Health & Life Sciences
An Alzheimer’s MRI model scored 94.5%. On new data, it fell to 61.6%
For a patient, the important test is whether a model works beyond familiar images. This classifier’s accuracy dropped sharply on an external dataset, and it remains a research system rather than a clinical diagnostic tool.
7 min · 1 source
Health & Life Sciences
Can brain MRI reveal more than BMI?
A peer-reviewed study trained a deep-learning model on 45,702 MRI scans from six cohorts. Brain-derived features tracked BMI and separated several disease groups better than BMI alone—but the disease analysis stayed inside UK Biobank and cannot establish cause, diagnosis or clinical benefit.
9 min · 2 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.