A commercially available AI device for detecting appendicular fractures showed consistent performance across three European hospitals and improved clinician sensitivity, according to a study published August 18 2026 in Radiology.
Led by a group at Erasmus MC University Medical Center in Rotterdam, the Netherlands, researchers evaluated RBfracture (Radiobotics, Copenhagen, Denmark), a CE-marked device for detecting extremity fractures on x-rays, with findings suggesting the tool generalizes robustly across health care systems and can benefit less experienced readers.
“Unlike curated or case-control datasets, we used consecutive cases, offering a more realistic assessment of the performance of the AI tool by capturing real-world conditions,” noted lead author Huibert Ruitenbeek, PhD, and colleagues. “Additionally, we validated the AI tool across datasets from different countries.”
Emergency department (ED) visits for acute fractures have been increasing globally, driven in part by the aging of the population, according to the authors. This increasing burden has placed significant demands on health care professionals and medical resources, they added. AI fracture detection tools have shown promise in improving diagnostic efficiency, but their performance across countries and health care settings and their effects on reader performance remain underexplored, the group wrote.
To approximate a “real-world” test of a device in this setting, the researchers use it to analyze x-rays of patients aged 21 or older with suspected appendicular fractures at academic medical centers in Denmark, Germany, and the Netherlands (n = 500 per hospital), collected consecutively between April 2018 to July 2022. In the multireader component of the study, 12 clinicians (four per site, spanning senior and junior radiologists and emergency physicians) evaluated 500 cases from their own institution with and without AI assistance.
Artificial intelligence (AI) assistance facilitates correct fracture identification in cases initially missed by human readers. A representative case is shown using anteroposterior radiographs in a 29-year-old male patient with wrist trauma (suspected nondisplaced transverse scaphoid fracture). The reader’s initial unaided assessment (top) resulted in a missed fracture, whereas with AI assistance (bottom), the fracture was correctly identified. The green box represents the fracture location as predicted by the AI tool, the blue box represents the reference standard annotation, and the red dots represent individual reader annotations. RSNAAccording to the findings, the hospital-level sensitivities of the AI tool ranged from 400 of 500 (80%) to 430 of 500 (86%) and specificities ranged from 465 of 500 (93%) to 480 of 500 (96%). An anatomy-based subgroup analysis revealed lower sensitivity of 105 of 140 (75%) for wrist/hand and finger examinations in one hospital versus 125 of 138 (91%) and 150 of 160 (94%) (p = .004) at the two others, but no differences in overall sensitivity or specificity were observed across all hospitals.
Lastly, use of the AI tool improved overall reader sensitivity (11 percentage points; p
“No differences in AI performance were observed across centers, and AI assistance was associated with higher reader sensitivity,” the authors wrote.
Next steps will be to evaluate the device’s performance before and after local calibration and to develop frameworks for continuous performance monitoring, the authors wrote. Prospective validation of workflow efficiency and patient outcomes in emergency settings will also be needed to support broader clinical implementation, the group concluded.
The full study is available here.