Can AI Trained on Adults Help Care for Newborn Babies?
When a newborn baby needs help breathing, the position of the breathing tube matters. The tube must sit correctly within a very small airway. If it extends too deeply, it can direct ventilation mainly into one lung.
Clinicians use chest X-rays to assess the tube’s position. Could artificial intelligence help them identify images that need urgent attention?
An international study that I had the honor to co-authored explored this question by adapting an AI model developed for adult chest X-rays to newborn care. The findings offer an encouraging example of how collaboration across hospitals could help make medical AI more useful for our smallest patients.
Why adult AI needs to learn about babies
AI systems learn from examples. A model trained on adult X-rays gains experience with adult anatomy and the imaging practices used in adult care. Newborn babies have much smaller structures, different body proportions, and their own clinical challenges.
We therefore cannot assume that an AI tool that performs well in adults will work equally well in newborns.
Developing a new model specifically for babies also presents challenges. Collecting and carefully labelling large numbers of neonatal images takes time, and individual hospitals may have relatively small datasets.
Our study examined whether hospitals could build on an existing adult model by helping it learn from neonatal images across several centres.
Learning from 30 neonatal intensive care units
The study included 2,965 chest X-rays from 2,960 infants across 30 neonatal intensive care units in 20 countries and territories.
The model’s task was specific: identify the tip of the breathing tube and the carina, the point where the windpipe divides into the two main airways leading to the lungs. Their relationship can help identify tubes positioned too deeply.
Researchers compared the original adult model with a model adapted using images from one hospital and a model trained sequentially using images from multiple hospitals.
This sequential approach, called continual learning, allows each training step to build on the experience gained at previous hospitals. Each centre contributed 50 training images.
What did we find?
Continual learning improved the model’s ability to locate the tube tip and carina compared with adaptation at a single hospital.
Improvements occurred at 29 of the 30 hospitals. The approach also improved the ranking of images showing potentially deep breathing tubes at 27 of 30 hospitals.
These findings suggest that learning from several hospitals can help an adult AI model adapt to the diversity of neonatal images. They do not mean that the model was correct in every case.
Could hospitals collaborate while keeping images locally?
The proposed approach would allow hospitals to exchange updated model parameters while keeping patient images within their own systems.
That is an important potential advantage. However, the experiments in this study ran centrally to simulate the hospital-to-hospital learning process. Testing the approach on separate hospital systems remains an essential next step.
What could this mean for babies and families?
A possible future role is to flag concerning X-rays for faster clinician review. The clinician would still interpret the image alongside the baby’s condition and make treatment decisions.
This study evaluated existing X-rays. It did not establish that using the model leads to faster tube correction, fewer complications, or better outcomes for babies.
Those are the questions that matter most as this work moves forward. Prospective studies must evaluate performance in everyday care, including false alerts and reliability in the smallest preterm infants.
The findings are promising: an adult AI model can become more useful for newborn imaging through learning across hospitals. The next challenge is to demonstrate that this improvement translates into better care.
Disclosure: I am a co-author of this study.
Study: Rao VM and colleagues. Repurposing adult AI radiograph models for neonatal care through continual learning: an international multicentre study. Nature Biomedical Engineering. Read the paper.
Comments