救急治療室における AI 支援による骨折検出の評価 (FAIR)
整形外科救急病棟における AI 支援骨折検出のコスト効率とワークフローへの影響を評価する
This study evaluates whether artificial intelligence (AI) can support doctors who interpret X-rays for suspected fractures in emergency care.
In the participating hospitals, X-rays are usually interpreted first by the frontline treating physician, while the formal radiology report is generally available later and not before the patient leaves the emergency department. AI may therefore provide an immediate additional assessment while clinical decisions are being made.
Patients were randomly assigned to one of two groups. In the AI-assisted group, physicians interpreted the X-rays with support from an AI system. In the control group, physicians interpreted the same type of X-rays without AI support. All final diagnoses and treatment decisions remained with the treating physician.
The main question is whether AI assistance affects the time from emergency department triage to discharge. The study also evaluates whether AI influences physician diagnostic confidence, the use of additional imaging, missed fractures, and diagnostic accuracy.
The study included patients aged 2 years or older presenting with an isolated musculoskeletal injury or relevant joint complaint requiring X-ray imaging. No additional imaging or treatment was required solely because of study participation.
調査の概要
詳細な説明
The FAIR (Fracture detection with AI in emergency Radiography) Trial was a prospective, international, multicentre, pragmatic randomised controlled trial evaluating the clinical impact of artificial intelligence (AI)-assisted interpretation of musculoskeletal radiographs in routine emergency care. The study was conducted across four clinical units at three hospitals in Austria and Germany.
Clinical workflow and randomisation
In the participating departments, radiographs were interpreted by frontline orthopaedic trauma or paediatric physicians, who made immediate diagnostic and treatment decisions. Formal radiology reports were generally available later. The intervention therefore evaluated AI as an additional decision-support tool rather than as a replacement for physician interpretation. Patient encounters were individually randomised using a single sequence generated by the trial statistician. Allocation was concealed using folded, sequentially numbered study forms and revealed after radiography. Patients were then directed to designated AI-assisted or control treatment areas. All clinical decisions remained the responsibility of the treating physician.
AI implementation
The intervention used BoneView version 2.3.8 (Gleamer, Paris, France), which provided visual annotations and classifications for fractures, focal bone lesions, dislocations, and joint effusions. The software version remained unchanged throughout recruitment. At Salzburg and Hallein, BoneView was delivered through Aidoc aiOS version 3.24.0; at Nuremberg, it was integrated through Visage version 7.1.19. Participating physicians received standardised training before using the system. AI output was documented on study case-report forms, although detailed software telemetry on actual viewing behaviour was unavailable.
Protocol evolution
During recruitment, an exploratory unblinded analysis in October 2025 examined prospectively collected clinical and workflow outcomes. Following methodological feedback during the Austrian Science Fund review process in March 2026, the outcome hierarchy for the fracture-focused clinical-effectiveness analysis was revised in consultation with the trial statistician, prioritising time from emergency department triage to discharge over diagnostic accuracy. The amendment did not alter randomisation, clinical procedures, or data collection. The originally registered cost-efficiency outcome was discontinued because the available data did not permit a reliable health-economic evaluation.
Reference standard
For fracture-related diagnostic analyses, a radiologist experienced in musculoskeletal imaging and a senior orthopaedic trauma surgeon reviewed the index radiographs and available clinical information. Disagreements were resolved by consensus or, when necessary, by a third expert. The assessment distinguished fractures visible on the original radiographs from fractures detectable only on subsequent imaging.
研究の種類
入学 (実際)
段階
- 適用できない
連絡先と場所
参加基準
適格基準
就学可能な年齢
健康ボランティアの受け入れ
受講資格のある性別
説明
Inclusion Criteria:
- Age 2 years or older.
- Presentation to the emergency department following trauma or relevant joint complaint requiring plain radiographic imaging.
- Injury involving a single anatomical region within the supported anatomical scope of the AI system.
- Written informed consent provided by the participant or legally authorized representative, with age-appropriate assent where applicable.
Exclusion Criteria:
- Age younger than 2 years.
- Injuries involving multiple anatomical regions.
- Head or cervical spine injuries.
- Previous radiographic imaging or medical assessment for the same injury before the index presentation.
- Contraindication to X-ray imaging, including pregnancy.
- Lack of informed consent.
Reduced image quality was not an exclusion criterion. Multiple fractures within the same anatomical injury region were eligible.
研究計画
研究はどのように設計されていますか？
デザインの詳細
- 主な目的：診断
- 割り当て：ランダム化
- 介入モデル：並列代入
- マスキング：なし（オープンラベル）
武器と介入
参加者グループ / アーム |
介入・治療 |
|---|---|
|
介入なし：Standard Radiograph Interpretation Without AI
Plain radiographs are interpreted by the frontline treating physician according to routine clinical practice without access to AI output. All diagnostic, imaging, treatment, consultation, and discharge decisions remain the responsibility of the treating physician. Formal radiology reports are generally available later and are not routinely available before completion of the emergency department encounter. |
|
|
実験的：AI-Assisted Radiograph Interpretation
Plain radiographs are interpreted by the frontline treating physician with access to real-time output from BoneView version 2.3.8 (Gleamer, Paris, France). The AI output is used as diagnostic decision support only; all final diagnostic, imaging, treatment, consultation, and discharge decisions remain the responsibility of the treating physician. |
BoneView version 2.3.8 (Gleamer, Paris, France) analyzes DICOM radiographs and provides real-time visual annotations and classifications for supported musculoskeletal abnormalities. Physicians in the intervention group could view fracture-related findings as well as other supported outputs, including dislocations, joint effusions, and focal bone lesions. At University Hospital Salzburg and Regional Hospital Hallein, BoneView was delivered through the Aidoc aiOS platform (version 3.24.0). At University Hospital Nuremberg, BoneView was integrated directly into the local imaging workflow. BoneView was used as a decision-support tool and did not replace physician interpretation or the subsequent formal radiology report. |
この研究は何を測定していますか？
主要な結果の測定
結果測定 |
メジャーの説明 |
時間枠 |
|---|---|---|
|
Time from Emergency Department Triage to Discharge
時間枠：During the index emergency department encounter (day 1).
|
Time in minutes from formal emergency department triage to discharge during the index encounter, compared between the AI-assisted and control groups. |
During the index emergency department encounter (day 1).
|
二次結果の測定
結果測定 |
メジャーの説明 |
時間枠 |
|---|---|---|
|
Missed fractures
時間枠：At the initial diagnostic assessment during the index emergency department encounter (day 1), with subsequent expert reference-standard adjudication.
|
Proportion of patient encounters with a fracture visible on the index radiographs, according to the expert-adjudicated reference standard, in which the fracture was not identified by the frontline treating physician during the initial diagnostic assessment. The reference standard is established by a radiologist and a senior orthopaedic trauma surgeon, with disagreements resolved by consensus or a third expert. Fractures identified on subsequent imaging but not visible on the original radiographs are not classified as missed radiographically visible fractures. |
At the initial diagnostic assessment during the index emergency department encounter (day 1), with subsequent expert reference-standard adjudication.
|
|
Additional Diagnostic Imaging Requested
時間枠：During the index emergency department encounter (day 1), following initial radiographic assessment.
|
Proportion of patient encounters in which the treating physician requested additional diagnostic imaging during the index emergency department visit. The frequency of additional imaging requests is compared between the AI-assisted and control groups. |
During the index emergency department encounter (day 1), following initial radiographic assessment.
|
|
Physician Diagnostic Confidence
時間枠：Measured immediately after the diagnosis
|
The level of confidence reported by physicians in their diagnostic decisions, measured on a Likert scale (1-10). This will compare how confident physicians feel when using AI assistance versus relying solely on their expertise. |
Measured immediately after the diagnosis
|
|
Diagnostic Accuracy of Fracture/Dislocation/Effusion/Bone Lesion Detection
時間枠：At the time of initial diagnosis, within 2 hours of patient presentation to the orthopedic emergency unit
|
The primary outcome measures the diagnostic accuracy of detecting broken bones/dislocations/effusions/bone lesions using sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). Diagnostic accuracy will be compared between the AI-assisted diagnostic approach and the standard physician-only approach. The gold standard for comparison will be determined by expert consensus based on independent review by a radiologist and an orthopedic specialist. |
At the time of initial diagnosis, within 2 hours of patient presentation to the orthopedic emergency unit
|
協力者と研究者
出版物と役立つリンク
研究記録日
主要日程の研究
研究開始 (実際)
一次修了 (実際)
研究の完了 (実際)
試験登録日
最初に提出
QC基準を満たした最初の提出物
最初の投稿 (実際)
学習記録の更新
投稿された最後の更新 (実際)
QC基準を満たした最後の更新が送信されました
最終確認日
詳しくは
本研究に関する用語
その他の研究ID番号
- KI-FRACTURE_001_2024-11-27
個々の参加者データ (IPD) の計画
個々の参加者データ (IPD) を共有する予定はありますか?
IPD プランの説明
De-identified individual participant data underlying the results reported in the primary publication will be made available to qualified researchers upon reasonable written request.
Shared data may include participant demographics, anatomical injury region, frontline physician diagnosis, AI findings, expert-adjudicated reference-standard findings, diagnostic confidence, additional imaging, and timing data used to derive the primary endpoint.
Data will be shared only for scientifically justified research purposes following review and approval of the proposed project by the study investigators and execution of an appropriate data-sharing agreement. Data sharing will comply with applicable institutional requirements, the General Data Protection Regulation (GDPR), and relevant Austrian and German data-protection legislation.
IPD 共有時間枠
Beginning 3 months after publication of the final prespecified FAIR outcome analysis, for 5 years
IPD 共有アクセス基準
Scientifically sound proposal; investigator review; applicable consent, ethics, institutional and data-protection requirements; appropriate data-sharing agreement
IPD 共有サポート情報タイプ
- STUDY_PROTOCOL
- SAP
医薬品およびデバイス情報、研究文書
米国FDA規制医薬品の研究
米国FDA規制機器製品の研究
米国で製造され、米国から輸出された製品。
この情報は、Web サイト clinicaltrials.gov から変更なしで直接取得したものです。研究の詳細を変更、削除、または更新するリクエストがある場合は、register@clinicaltrials.gov。 までご連絡ください。 clinicaltrials.gov に変更が加えられるとすぐに、ウェブサイトでも自動的に更新されます。