Text

Objectives and questions: Artificial intelligence (AI) is increasingly being integrated into traumatology, particularly for assisting in fracture diagnosis. Over the past few years, AI-based fracture detection software has undergone extensive evaluation, with retrospective studies suggesting improvements in detection accuracy. However, its real-world impact on clinical decision-making, workflow efficiency, and resource utilization in emergency settings remains uncertain. This study aims to assess the unintended consequences and hidden costs of AI-assisted fracture detection, with a particular focus on its effects on additional imaging rates, physician uncertainty, and patient flow in high-pressure emergency environments.

Material and methods: This prospective, multicenter, randomized controlled trial is conducted at trauma centers in university hospitals across two countries, concluding recruitment on May 30, 2025. Patients of all ages with suspected limb fractures are randomized (1:1) using computer-generated concealed allocation into either an AI-assisted diagnostic workflow or standard physician-only interpretation. AI support is provided via an FDA- and CE-approved deep learning model, displaying results as “Yes”, “Doubt”, or “No”. The blinded expert panel (two radiologists, one orthopedic surgeon) establishes the reference standard. Primary outcomes include additional imaging, physician uncertainty, waiting times, and ED treatment duration. Secondary outcomes assess diagnostic accuracy, decision-making impact, and economic implications.

Results: Preliminary analyses indicate that AI assistance may significantly impact physician decision-making, potentially increasing additional imaging requests and prolonging ED treatment times due to uncertainty surrounding AI-generated findings. While AI is expected to enhance sensitivity in fracture detection, its effect on specificity and false-positive rates remains a key consideration. The trial aims to assess whether AI integration improves diagnostic confidence among emergency physicians or introduces workflow inefficiencies that could offset its potential benefits. Additionally, cost-effectiveness analyses will evaluate whether reductions in missed fractures justify the increased financial burden associated with AI implementation, considering its impact on resource utilization and hospital expenditures.

Discussion and conclusions: This study is the first multicenter, randomized controlled trial to evaluate AI-assisted fracture detection in a real-world emergency setting. By assessing AI’s impact on workflow efficiency, physician uncertainty, and healthcare resource utilization, the findings will provide critical insights into both the benefits and limitations of AI integration in emergency medicine. The results will inform an evidence-based approach for incorporating AI into clinical practice, addressing key challenges related to diagnostic accuracy, cost-effectiveness, and operational efficiency. Ultimately, this study will help determine whether AI enhances emergency fracture diagnosis or introduces unintended workflow inefficiencies that may offset its potential advantages.