Design: Systematic review and diagnostic accuracy meta-analysis
Population: 28 studies (17 pooled) of AI-assisted versus unassisted fracture reading
Primary outcome: Sensitivity and specificity for fracture detection
Effect: Sensitivity 87% vs 73%; specificity 95% vs 94%; junior clinicians +21 points

This meta-analysis included 28 studies, 17 pooled, comparing clinicians reading fracture imaging with and without AI assistance.

With AI, pooled sensitivity rose from 73% (95% CI 69% to 78%) to 87% (84% to 89%), while specificity was unchanged at about 95%. The summary area under the curve rose from 0.85 to 0.93. Junior clinicians gained most, an absolute increase in sensitivity of 21 percentage points. Plain radiographs and junior status still limited the best accuracy achievable.

The studies were mostly reader studies, not trials of patient outcomes. AI support looks most useful where junior doctors read their own films out of hours.

Consider AI fracture detection support where junior doctors read their own radiographs.
Keep a low threshold for CT when clinical suspicion stays high despite a normal radiograph, with or without AI.
Treat an AI flag as a prompt to look again, not a diagnosis.
Audit missed fractures before and after introducing any AI tool.

Why it matters

Missed fractures are among the commonest emergency department errors, and most are made by the least experienced readers.

The statistics, in plain English

Sensitivity of 87% means about 13 in 100 fractures would still be missed with AI help, against about 27 without it. Reader studies can overstate real-world gains, because readers know they are being tested.