AI raises fracture detection sensitivity by 14 points, and most for juniors

Design

systematic review and meta-analysis following PRISMA-DTA, bivariate random-effects model, QUADAS-3 quality assessment

Population

28 included studies, 17 pooled, comparing AI-assisted with independent physician fracture interpretation

Primary outcome

pooled sensitivity and specificity for fracture detection

Effect

sensitivity 87% (95% CI 84–89) assisted vs 73% (69–78) unassisted; specificity 95% (92–97) vs 94% (89–96); summary AUC 0.929 vs 0.849; junior clinicians gained 21 absolute points of sensitivity

Missed fractures are among the commonest diagnostic errors in emergency departments, and they concentrate in films read overnight by the least experienced clinician available. This meta-analysis pooled 17 studies, from 28 meeting inclusion criteria, comparing AI-assisted with independent physician interpretation for fracture detection, using a bivariate random-effects model.

Pooled sensitivity rose from 73% (95% CI 69–78) unassisted to 87% (84–89) with AI assistance, while specificity held at about 95% (92–97) against 94% (89–96). The summary area under the curve moved from 0.849 to 0.929. The gain was not evenly distributed: junior clinicians gained 21 absolute percentage points of sensitivity, and meta-regression identified junior status and two-dimensional radiography as independent predictors of where the assisted ceiling sat.

The authors' own conclusion is the honest one and it is worth repeating: AI assistance narrows the experience gap but does not abolish the ceiling set by clinical expertise and by the physical limits of a plain film. A scaphoid or an occult hip fracture that is not on the radiograph will not be on the radiograph with an algorithm looking at it. The threshold for cross-sectional imaging should not move.

• Read the film yourself first, then look at the AI output — assistance studies measure a clinician using a tool, not a tool working alone.
• The benefit is largest overnight and in units staffed by trainees; that is where deployment pays.
• Do not let a negative AI result raise your threshold for CT or MRI in a clinically convincing occult fracture.
• Specificity was maintained, so this was not bought with a flood of false positives — but that holds at study-level thresholds, which may not match the ones your vendor ships.
• Most of these studies assessed accuracy on stored images, not the clinical decision that followed.

Why it matters

It quantifies where the tool helps — the inexperienced reader on a plain film — rather than treating AI assistance as a uniform upgrade.

Don't overread it

These are diagnostic accuracy studies on images, not trials showing that patients had better outcomes.

The statistics, in plain English

A sensitivity rise from 73% to 87% means roughly half the previously missed fractures are caught — a meaningful shift, and the confidence intervals do not overlap. Specificity staying at 94 to 95% is the critical control: an easy way to raise sensitivity is to call more things fractures, and that did not happen here. Meta-regression explained 52.2% of the heterogeneity, which leaves nearly half unexplained, and only 17 of 28 eligible studies could be pooled — both signs that the studies differed in ways the analysis could not fully account for.

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.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.