Artificial intelligence in virtual fracture clinics: a systematic review of imaging and clinical-text tools

TL;DR

This review examines AI tools for virtual fracture clinics, finding high performance in imaging tools but a lack of NLP tools for clinical documentation.

Contribution

The study is the first to systematically review AI tools combining imaging and clinical-text for virtual fracture clinics.

Findings

Commercial AI tools for fracture detection showed high sensitivity, especially for wrist-specific models.

Researcher-developed models often outperformed commercial tools in sensitivity.

No NLP tools were found for acute orthopedic care, highlighting a gap in AI-VFC integration.

Abstract

Virtual fracture clinics (VFCs) are a well-established component of acute orthopedic management pathways. Artificial intelligence (AI) healthcare tools are increasingly sophisticated and have the potential to disrupt current practices. The aim of this review was to determine the opportunities, performance and readiness of AI systems that integrate clinical-text and imaging data for the triage or management of patients in VFCs.

A search of MEDLINE and Embase was performed between January 2010 and July 2025. The review included primary research studies investigating AI for fracture detection via X-rays and natural language processing (NLP) models for clinical documentation. A random-effects meta-analysis was conducted to calculate pooled sensitivity and specificity, stratified by anatomical region and developer type (commercial vs. researcher-developed).

We included 54 studies: 52 imaging/X-ray studies and 2 NLP/clinical-text studies. Among the imaging studies, 13 evaluated commercial tools, and 39 evaluated researcher-developed models. There were 2 NLP models, both of which interpreted radiology reports rather than text summaries of clinical assessments. No studies that included the use of NLP models in acute orthopedic care could be found. A meta-analysis of commercial tools (n = 11) demonstrated a pooled sensitivity across both multiregional “Limb” tools of 92.58% (95% CI 90.61–94.17%) and anatomy-specific “Wrist” tools of 89.95% (95% CI 72.18–96.86%). Wrist-specific commercial tools demonstrated higher specificity (96.80%; 95% CI 90.12–99.01%) compared to general limb tools (89.69%; 95% CI 84.02–93.51%), suggesting that anatomical targeting may reduce the number of false positives. Researcher-developed models (n = 32) often reported superior metrics for sensitivity compared to the sensitivity of commercial tools.

VFCs require the integration of information from imaging and patient records. Multiple image interpretation tools are available with high performance in fracture identification. The development and integration of NLP tools to interpret clinical documentation from emergency departments and urgent care centers are necessary for AI-VFC.

The online version contains supplementary material available at 10.1186/s13018-025-06656-5.

Figures