Truace
Subscribe

Generalisability of an adult-trained AI model to paediatric coeliac disease diagnosis: a multicentre external validation study

Background: Coeliac disease diagnosis on duodenal biopsies is limited by interobserver variability. We have previously demonstrated pathologist-level performance with our artificial intelligence (AI) model for the histopathological diagnosis of adult coeliac disease, but not in paediatric practice. As paediatric coeliac disease screening programmes expand internationally, accurate and scalable diagnostic tools are needed. We investigated whether an AI model trained exclusively on adult whole-slide images (WSIs)…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-2186· cite & verify
Generalisability of an adult-trained AI model to paediatric coeliac disease diagnosis: a multicentre external validation study
Hospital Universitari Doctor Peset, València 04 by 19Tarrestnom65. CC BY-SA 4.0 · https://creativecommons.org/licenses/by-sa/4.0

In brief

Researchers trained an ensemble of multiple-instance learning models on 9958 whole-slide images from 8421 adults across five centres, then tested it on 708 consecutive children from Edinburgh and Southampton, which were not part of training. In the children the model reached an AUC of 98.8% for coeliac disease.

The result suggests adult-derived pathology AI could help meet growing demand as paediatric screening programmes expand, by supporting confident identification of negative biopsies. The study is a retrospective external validation, and sensitivity in children was lower than in adults, so how it performs as a clinical aid remains to be shown.

Main points

  1. Adult cross-validation gave an AUC of 99.1% and sensitivity of 89.3%, versus 98.8% AUC and 80.2% sensitivity in the paediatric test set.
  2. Excluding intermediate-probability predictions (604 of 708 cases) raised sensitivity to 100% and specificity to 98.7%.
  3. The 467 high-confidence cases (66.0%) had no misclassifications, and expected calibration error was 0.03.
  4. Performance improved significantly when biopsies from both the duodenal bulb and the descending duodenum were considered.

The gain

An AI model trained only on adult biopsy slides kept high accuracy on 708 children from two untrained-on centres, and could flag coeliac-negative biopsies within defined confidence thresholds.

The rundown

The training and validation set covered 961 adults with coeliac disease, with samples collected from January 2014 to March 2025. The model used features from a foundation model feeding an ensemble of multiple-instance learning classifiers.

The paediatric test cohort included 86 children with coeliac disease out of 708. Paediatric specificity was 98.4% and negative predictive value 97.3%.

Probability outputs were grouped into clinically interpretable categories. Cases below 10% or at or above 65% predicted probability were treated as confident calls, while those in between were set aside as intermediate. The authors say this could support safe identification of negative biopsies within defined thresholds. The work was funded by the National Institute for Health and Care Research.

Sources

  1. Peer-reviewedeClinicalMedicine2026-10-12

The debate