- Design
- Systematic review and bivariate random-effects meta-analysis, PRISMA-DTA, QUADAS-3 quality assessment
- Population
- 28 studies included, 17 pooled, comparing AI-assisted with independent physician fracture interpretation
- Primary outcome
- Pooled sensitivity and specificity for fracture detection, and summary ROC area under the curve
- Effect
- Sensitivity 87% (95% CI 84-89) assisted versus 73% (69-78) unassisted; specificity 95% (92-97) versus 94% (89-96); SROC-AUC 0.929 versus 0.849; junior clinicians +21% absolute sensitivity
This systematic review and meta-analysis followed PRISMA-DTA, searched PubMed and Web of Science to September 2025, assessed quality with QUADAS-3, and pooled 17 of 28 included studies with a bivariate random-effects model.
AI assistance raised pooled sensitivity from 73% (95% CI 69-78) to 87% (95% CI 84-89), with specificity essentially unchanged at 95% (92-97) versus 94% (89-96). Summary ROC area under the curve went from 0.849 to 0.929. That is the important shape of the result: the gain came in missed fractures, not at the cost of false positives.
The subgroup finding is where the practical value is. Junior clinicians gained a 21% absolute increase in sensitivity — the tool closes an experience gap rather than adding to what an experienced reader already sees. Multivariate meta-regression, explaining 52.2% of heterogeneity, identified two-dimensional radiography and junior physician status as independent predictors of a lower assisted ceiling.
That last point is the one worth carrying. AI does not overcome the limits of a plain film. A scaphoid or an occult hip fracture that is not visible on the radiograph does not become visible because software looked at it. The threshold for cross-sectional imaging in a patient with convincing clinical findings and a negative assisted read should not move.
- Treat an assisted negative read the same as an unassisted one when clinical suspicion is high
- The benefit is concentrated in less experienced readers — deploy it where they work
- Specificity did not fall, so the workload cost of false positives should be small
- Two-dimensional radiography sets a ceiling no algorithm crosses
- Ask what the tool was validated on before trusting it on a body region it did not see in training
Why it matters
It reframes AI fracture tools as a fix for an experience gap rather than a general accuracy upgrade.
Don't overread it
These are pooled study-level accuracy figures, not evidence that deploying a tool improves patient outcomes in a real department.
The statistics, in plain English
A sensitivity rise from 73% to 87% means missed fractures fell from about 27 in 100 to about 13 in 100 in these study populations — a halving, though on a base rate that reflects trial datasets rather than your department. Specificity staying at 94-95% matters just as much: a sensitivity gain bought with false positives would simply move the work elsewhere. Meta-regression explaining 52.2% of heterogeneity means about half the variation between studies is still unexplained.
Read the rest in the app
You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.
Scan to keep reading on your phone. No account needed to start.
Tomorrow morning, before your first patient
One edition a day for orthopaedics, written by the desk, every claim tied to its paper. Six minutes.
Get the app — free