- Design
- Systematic review and diagnostic meta-analysis
- Population
- 17 studies of clinicians interpreting fracture imaging with and without AI
- Primary outcome
- Pooled sensitivity and specificity for fracture detection
- Effect
- Sensitivity 87% vs 73%; specificity 95% vs 94%; AUC 0.929 vs 0.849
This diagnostic meta-analysis pooled 17 studies comparing clinicians reading fracture imaging with and without AI assistance.
AI assistance raised pooled sensitivity from 73% (95% CI 69 to 78%) to 87% (84 to 89%) with specificity unchanged (94% to 95%). Area under the summary ROC curve rose from 0.849 to 0.929. Junior clinicians gained most, with an absolute sensitivity increase of 21 points. Plain radiographs and junior status still predicted a lower ceiling of accuracy even with AI.
For orthopaedic departments, the main effect is fewer missed fractures in the emergency department and fracture clinic, where junior doctors often make the first read. The authors are clear that AI does not remove the limits of two-dimensional radiographs, so a low threshold for CT or MRI when clinical suspicion is high still applies. The included studies are mostly reader studies, not trials measuring patient outcomes.
- Where AI fracture detection is available, use it as a second reader, especially for junior staff and out-of-hours reads.
- Keep a low threshold for CT or MRI when examination suggests a fracture the radiograph does not show, with or without AI.
- Review AI-flagged findings clinically; specificity was high but not perfect.
- Audit missed-fracture returns before and after adopting AI to see whether the gain holds locally.
Why it matters
Missed fractures on first read are a leading cause of complaints and late treatment.
Don't overread it
These are accuracy studies; none shows that patient outcomes improved.
The statistics, in plain English
Sensitivity is the share of real fractures detected; 87% vs 73% means about 14 more of every 100 fractures picked up. Specificity is the share of normal studies correctly called normal, and it stayed at about 95%, so AI did not add false positives.
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