Published in *Nature Biomedical Engineering*, the method locates the intubation probe with an average error of 0.67 cm, compared with 4.30 cm for the original adult model.
Children represent less than 1% of public medical imaging datasets, and only 17% of FDA-cleared medical AI devices are indicated for use in pediatrics. This is mainly due to highly protected neonatal data, which remains in the hospital where the child is being treated. A team from Harvard Medical School, led by Pranav Rajpurkar and Vishwanatha Rao (University of Pennsylvania), published a way to get around the obstacle on October 13, 2026.
A model who travels, not the images
Rather than centralizing the X-rays, the model is trained successively in each facility. The researchers adapted a model designed for adult chest X-rays from 2,965 images from 30 neonatal intensive care units in 20 countries and territories. To locate the tip of the intubation tube, the average error drops to 0.67 cm, compared to 1.29 cm for a model adapted in a single hospital and 4.30 cm for the adult model. The shared model outperforms local versions in 29 of the 30 sites.
Read the full article on the original source
Go to the source →