Adult-Trained AI for Childhood Coeliac Disease
An adult-trained pathology AI achieved 80.2% sensitivity and 98.4% specificity in 708 children from two external centers. Results varied with confidence thresholds.
A UK research team tested whether AI trained exclusively on adult duodenal biopsy images could classify childhood coeliac disease. In eClinicalMedicine, they report retained discrimination in a pediatric population excluded from training.
Transfer from adult to pediatric images
The researchers developed the model using 9,958 whole-slide images from 8,421 adults at five centers. They did not use pediatric images for training.
External evaluation involved 708 consecutive children at Edinburgh and Southampton, neither of which contributed training data. Of these children, 86 had coeliac disease. The evaluation crossed age groups, centers and scanner platforms.
Performance in children
Sensitivity was 80.2% (95% CI 70.6 to 87.3) and specificity was 98.4% (97.1 to 99.1). Sensitivity measures detection among those with the disease; specificity measures correct negative classifications among those without it. Negative predictive value was 97.3%, the proportion of negative results that were truly disease-negative.
AUC was 0.988 (95% CI 0.982 to 0.994). AUC measures discrimination, with 0.5 representing chance and 1.0 perfect separation. A high AUC does not describe the miss rate at a particular threshold; the full-cohort sensitivity remains relevant.
Confidence-based selection
The team also restricted analysis to predictions below 10% or at least 65% probability. Among these 604 children, or 85.3% of the pediatric cohort, sensitivity was 100% and specificity was 98.7%.
No misclassifications occurred among 467 children with predictions below 2% or at least 85%. These results apply to selected predictions that exclude uncertain cases, not to the full cohort of 708 children.
Sampling and clinical scope
Performance improved significantly when the researchers considered biopsies from both the duodenal bulb and the descending duodenum. The authors emphasize the importance of multisite biopsy sampling.
The study evaluated existing pathology images. The public abstract does not report effects on missed diagnoses or working time after clinical implementation. Transfer to children without pediatric training and clinical usefulness require separate evaluation.
Research note: The UK National Institute for Health and Care Research funded the study. The public abstract does not provide detailed competing-interest disclosures or Japanese regulatory approval information.
Sources
- [1]
- [2]
Images: generated with GPT Image 2.5 (illustrative).