Sequential training improves newborn X-ray AI in 30-hospital study

An AI model trained on adult chest X-rays became substantially better at assessing newborn breathing-tube placement after learning from hospitals one after another. A study published October 13, 2026, in Nature Biomedical Engineering found lower anatomical-location errors at 29 of 30 sites. Its training approach transfers the model between institutions rather than pooling their patient images.

Small patients, scarce training data

The retrospective study covered 30 neonatal intensive care units across 20 countries and territories. It addressed a specific problem: checking whether an endotracheal tube, the breathing tube used to treat respiratory problems, sits correctly in a newborn’s chest X-ray.

There is little margin for error. A tube placed too deeply can ventilate only one lung and cause lung collapse. False alarms also carry risks: unnecessary attempts to insert the tube again can destabilize a patient. The researchers focused on detecting deep placements because of their clinical severity.

Adult radiology tools do not necessarily transfer well to newborns, whose physiology and imaging protocols differ. Building pediatric datasets is difficult, too. The paper notes that children represent less than 1% of public medical imaging datasets, while privacy restrictions and the work of labeling scans complicate collaboration between hospitals.

Pass the model, not the scans

The team used continual learning: a model learns from successive datasets without access to the data it encountered earlier. Here, an existing adult-trained model was retrained using 50 neonatal radiographs from the first hospital. Its updated parameters—the numerical settings learned during training—then became the starting point for the next hospital.

That sequence continued through the participating sites, allowing each hospital’s training to build on what earlier institutions had contributed. The proposed workflow keeps retraining within each hospital’s computing environment and transfers only model parameters. Harvard Medical School’s account describes the experiment as simulating a setting in which patient data never leaves the hospital.

Unlike federated learning, which repeatedly collects and averages updates from multiple sites through a central server, this method passes the model along once. The researchers also checked whether the final model retained or improved performance at hospitals encountered earlier. That tests whether successive retraining accumulates useful knowledge rather than simply fitting the latest institution.

0.67 cmSequential training

The continually trained model had the lowest mean localization error.

1.29 cmHospital-only adaptation

Adapting the adult model using only one hospital’s data produced greater error.

4.30 cmUnadapted adult model

Applying the adult-trained model directly to neonatal images produced the greatest error.

Better localization, a narrower clinical claim

The comparison measured how accurately each approach located the tube tip and carina, anatomical landmarks used to assess placement. Its two alternatives reflected choices available to individual hospitals: use the adult model unchanged, or adapt it using local data alone. Sequential training also improved detection of deep intubations at 27 of 30 sites.

The intended use is to flag concerning X-rays for radiologist review, not replace that review. The researchers suggest such triage could shorten the time to confirmation of dangerous placements. But this was a retrospective proof of concept: its measured gains concern image interpretation, rather than demonstrated improvements in patient care.

In the Harvard Medical School account, the researchers propose extending the approach to other adult-trained models and populations underrepresented in training data. That possibility remains broader than the result established here. This study’s evidence is specific to adapting an adult radiograph model for neonatal breathing-tube assessment.