Context Determines Value of AI Doubt Alerts in Fracture Detection: A Multi-Site Retrospective Analysis of Clinical Indication and Anatomy

Poster presented at CLINICCAI 2026, the Clinical Translation day of the 29th International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2026), Strasbourg, France. Paper 48. Question. A fracture-detection AI (BoneView, Gleamer) returns a positive, doubt or negative output on every processed radiograph without clinical context. Does the referral indication decide what a doubt output is worth? Methods. 879 doubt outputs from 2025, 22 outpatient centres of the 3R Swiss Imaging Network (60 radiologists, AI in routine use since 2021), one per examination, research consent. Referral indication and radiologist conclusion classified from text by a rule-based classifier validated against blinded double reading (150 readings, two radiologists). Results. 479 of 879 doubt outputs (54.5%) followed a referral without a fracture question. The radiologist concluded an acute finding after 19.5% of doubt outputs with a fracture question and 7.9% without (odds ratio adjusted for body region 2.97, 95% CI 1.88-4.67). Implications. Pass the referral indication to the AI; without a fracture question, keep the result on demand; ask vendors for indication-specific operating points. At 3R, AI results are displayed passively with the study, without workflow interruption. Retrospective, observational analysis. Disclosures. B.R. holds equity in Gleamer, the manufacturer of the BoneView tool evaluated in this study. Data extraction and analysis were performed independently by S.M., N.H. and O.N., who have no financial relationship with Gleamer. S.M. reports an ongoing paid services/consulting relationship with 3R Swiss Imaging Network via Medlogic; S.M. is not an employee of 3R. B.R. was blinded second reader for label validation.