The abstract is read at the publisher; the summary is JClub's.
Pooling adult and paediatric training data enables a merged EfficientNet-B0 to reach pneumonia AUC-ROC 0.951 (95% CI 0.932–0.967) on paediatric images, fixing cross-domain failure.
- Why it matters: Hospital-specific training often generalizes poorly to other patient groups, which is especially harmful in low-resource settings lacking large local datasets.
- What they did: They trained lightweight CPU models (EfficientNet-B0, ResNet-18, and baselines) on a 10,000-image NIH ChestX-ray14 subset, then tested without retraining on paediatric pneumonia data and NIH test data (n = 2,500), using bootstrap confidence intervals.
- The result: Cross-domain single-domain models collapsed (AUC-ROC 0.454 adult→paediatric and 0.473 paediatric→adult), while pooling improved paediatric AUC but not adult-only performance; Monte Carlo dropout referral increased answered-case accuracy from 0.889 to 0.948 on NIH, without domain-shift gains.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.