Radiographer fracture detection with and without AI assistance in prospective emergency department use

Artificial intelligence (AI) systems for fracture detection perform strongly in controlled settings, but prospective real-world evidence is limited. In a three-month deployment in one emergency department, thirteen radiographers in a radiographer-led screening programme independently interpreted 2496 consecutive musculoskeletal examinations (one radiographer per examination), reviewed output from a commercial deep-learning system, then recorded a final decision, giving 2496 paired reads against a radiologist reference standard. Standalone AI sensitivity was 75.6%, falling to 46.2% for subtle fractures, below the radiographers’ baseline sensitivity of 91.6%. Radiographers’ baseline accuracy was high (97.6%; specificity 99.0%) and improved modestly with AI assistance (98.0%; exact McNemar p = 0.0064). With AI assistance, radiographers corrected 11 errors and introduced 1, a net gain of 10 cases. Subgroup gains were directionally consistent but not significant after correction. Radiographers changed their decision in only 11 of 729 discordant cases, and most negative interactions were correct AI recommendations not adopted. No or incomplete AI output was returned for 8.8% of acquired examinations. AI assistance delivered a small, selective benefit arising from complementary rather than superior algorithmic performance. Operational reliability proved as consequential as diagnostic accuracy, and behavioural inferences rest on few events. Downstream clinical outcomes were not measured.