How AI Keeps Getting Better Across Hospitals Without Sharing Data

International researchers ran a retrospective study to see if continual learning could help an AI model that checks endotracheal tube placement on chest X‑rays. They tested three ways to use the model: keep it as‑is, fine‑tune it at each new hospital, or let it learn continuously across sites. The dataset included 2,313 intensive‑care chest radiographs from 23 hospitals in 12 countries across five continents. One hospital was set aside for external validation. Patients were adults (≥18) who had tubes placed in 2021 or earlier when there were too few recent cases. The team gathered cases through keyword searches in hospital systems and cross‑checked them with radiology reports.

Each hospital contributed roughly 100 anteroposterior images, of which 50 were used for training. The images were stripped of any identifying details before being sent to the study team. Two radiologists independently reviewed and annotated every X‑ray to ensure consistency and accuracy.

The continual learning models recorded an average ETT‑carina error of 10.58 mm (standard deviation 12.98 mm). This beat the fine‑tuned‑only approach, which had an average error of 12.49 mm (SD 15.37 mm), and the original model, which averaged 16.39 mm (SD 21.56 mm). The continual learning versions outperformed the original model at every hospital and beat the fine‑tuned models in 21 of the 22 sites.

Even with just 50 training examples per hospital, continual learning proved superior to traditional fine‑tuning for generalizing the AI across diverse clinics. The findings suggest that continual learning can improve medical AI performance while keeping patient data private and avoiding complex computational setups.