The Garden I–II fracture that the film does not show

Design

multicentre retrospective model development and external validation with a 10-reader comparison study

Population

2,576 adults with suspected hip trauma at four hospitals who had radiographs plus same-episode CT or magnetic resonance imaging

Primary outcome

sensitivity, specificity and area under the curve for femoral neck fracture detection, and effect on reader performance

Effect

pooled sensitivity 97.5%, specificity 98.8%, area under the curve 0.99; in radiograph-negative or indeterminate fractures 94.7% versus 86.2% for radiologists and 68.8% for emergency physicians

A multicentre retrospective study assembled 2,576 adults with suspected hip trauma across four hospitals who had pelvic or hip radiographs and CT or magnetic resonance imaging in the same episode, giving a reference standard. A two-stage deep learning model, OccuNet, was trained with contrastive pretraining on artefact-augmented copies of the same radiograph before a fracture detector was fine-tuned, then tested internally and on three external sets.

On the pooled test group of 1,766, sensitivity was 97.5% and specificity 98.8%, area under the curve 0.99 (95% CI 0.99 to 0.99). The finding that matters sits in the subgroup of 189 radiograph-negative or indeterminate fractures — the Garden I and II injuries that get sent home. There the model's sensitivity was 94.7% against 86.2% for five musculoskeletal radiologists (P<0.001) and 68.8% for five emergency medicine physicians (P<0.001). Used as an aid, it raised radiologist sensitivity from 93.7% to 97.2% and emergency physician sensitivity from 84.3% to 95.6% (both P<0.001), while cutting mean reading time by 14.9% and 18.9% respectively.

The 68.8% figure is the one to sit with. Nearly a third of occult femoral neck fractures were missed by the clinicians who see these patients first and decide whether they go home, and that is where the displacement, the avascular necrosis and the hemiarthroplasty come from. A tool that closes most of that gap is doing something narrow and valuable rather than something general.

The qualifier is the design. Every patient here had cross-sectional imaging, so the cohort is enriched for the diagnostically difficult; and readers worked on a retrospective set knowing they were in a study, not on a night shift. Sensitivity gains in a reader study are the best case, not the deployed case.

• Keep low-threshold cross-sectional imaging for the older patient who cannot weight-bear, whatever the radiograph and whatever a model says
• If you evaluate a fracture-detection tool, ask for its performance on radiograph-negative cases specifically — overall sensitivity hides that
• Note the gain was largest for non-radiologist readers; that is where such a tool would be deployed first
• Treat the reading-time reduction as a secondary benefit, not the reason to adopt
• Audit your own missed occult neck-of-femur fractures before and after any change — the local rate is the number that matters

Why it matters

It puts a number on how often the injury that most needs finding early is missed by the clinician who decides whether the patient goes home.

Don't overread it

This is a retrospective reader study in a cohort enriched for difficult cases — it does not show that deploying the model reduces missed fractures in practice.

The statistics, in plain English

An area under the curve of 0.99 on a case set where every patient had CT or magnetic resonance imaging is not the performance to expect in an unselected emergency department, where the prevalence of fracture is far lower and the positive predictive value falls accordingly. The comparison that survives that caveat is the relative one: the same 189 hard cases, read by the model and by the humans, with a 26 percentage point gap over emergency physicians. Reader sensitivity rising from 84.3% to 95.6% with assistance is a within-reader comparison, which is the right design, but it was done retrospectively with full attention on the task.

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.