Emergency Radiology: A Testing Ground for AI
Plain radiography remains the most frequently ordered imaging examination in emergency departments. Fractures, dislocations, joint effusions, pneumothoraces, and pleural effusions are all conditions whose initial diagnosis most often relies on a radiograph interpreted in the first instance by an emergency physician, before the radiologist provides the definitive report. In this high-pressure environment, where time-to-care is constrained and radiological coverage at night and on weekends is rarely guaranteed, diagnostic errors represent a genuine risk.
According to some estimates, missed fractures account for up to 80% of interpretation errors on bone radiographs in emergency settings. Added to this reality is the growing pressure of examination volumes, professional burnout among teams, and the inherent variability of human reading, particularly on call. It is in this context that artificial intelligence (AI) solutions are being progressively deployed in emergency radiology departments.
But beyond the performance figures that are advertised, what are these tools actually worth under the conditions of daily practice?
Solid Performance for Fracture Detection
Deep learning algorithms dedicated to fracture detection achieve high levels of sensitivity and specificity. The overall performance of these tools for detecting acute fractures falls around a sensitivity of 95 to 96% and a specificity of 97 to 98%, with negative predictive values exceeding 99%.
This high negative predictive value is particularly valuable in an emergency setting: a negative AI result constitutes a strong argument for ruling out a fracture and avoiding unnecessary immobilization, thereby contributing to reducing the radiologist’s workload and securing medical decision-making in the absence of immediate radiological coverage. This “exclusion rule” confers genuine value on this tool in an emergency context.
However, some software generates a significant number of indeterminate or false-positive results. This tendency toward over-flagging of uncertainties, while clinically less problematic than a false negative, requires vigilance when integrating these tools into the clinical workflow.
Detection of Dislocations and Joint Effusions: The Achilles Heel of Current Algorithms
While the performance of AI tools for fracture detection is well documented and broadly reassuring, their ability to identify other traumatic pathologies, such as joint dislocations and effusions, reveals significant gaps.
This insufficient performance is explained in part by the low prevalence of dislocations in real-world practice cohorts, making it difficult to train on a sufficient number of examples, and also by the morphological diversity of dislocations depending on the joint involved.
This asymmetry in performance across different pathologies is an important signal for industry: a tool marketed as a comprehensive detection solution for musculoskeletal emergencies cannot afford to perform well only for fractures. The clinical credibility of these systems depends on balanced coverage of the full pathological spectrum.
Beyond the Skeleton: Multi-Pathology AI as a Tool for Complete Emergency Flow Coverage
A major conceptual evolution in the deployment of AI in emergency radiology consists in moving beyond the “one algorithm, one pathology” logic toward solutions capable of handling the entire emergency radiograph flow, whether musculoskeletal or thoracic examinations, adults or children.
Multi-pathology algorithms that maintain consistent performance across age subgroups (pediatric, adult, geriatric), different anatomical regions, and both thoracic and musculoskeletal examinations are the most promising for clinical generalization.
The geriatric population deserves specific mention. Over-represented in emergency flows, it accumulates multiple diagnostic challenges: osteopenia making fractures more subtle, osteoarthritis obscuring bony contours, and polymorbidity.
A particularly striking result concerns cases of discordance between emergency physicians and radiologists. For cases that are difficult to interpret for an emergency physician alone, the AI algorithm achieves an accuracy of 90%, demonstrating that AI and emergency physicians have the prerequisites for effective diagnostic complementarity. In cases misdiagnosed by emergency physicians, AI would correct the diagnosis in 9 out of 10 cases. Conversely, in cases misclassified by AI, emergency physicians maintain an accuracy above 90%. These two actors do not share the same blind spots, which fully justifies their use in synergy.
AI as a Triage and Workflow Organization Tool in Radiology
Beyond diagnostic performance alone, it is the question of integrating AI into the organization of the radiological workflow that will largely determine its real impact on care.
The most promising model repositions AI not as a tool to replace the radiologist, but as a two-tiered prioritization and safety system. Upstream, the algorithm’s high negative predictive value allows identification of examinations raising no diagnostic concern; these cases can be reviewed by the radiologist on a deferred basis, without urgency. Downstream, cases flagged as positive or indeterminate by AI, as well as cases of radio-clinical discordance, are prioritized for immediate radiological review.
This intelligent triage model is all the more relevant given that 24/7 radiological coverage is structurally difficult to maintain in many institutions. In this context, AI does not fill a human void but intelligently organizes the workload by directing radiological expertise where it is most needed.
One point of attention concerns the “indeterminate cases” generated by AI. These uncertain results, which represent approximately 7 to 12% of the flow, should not be automatically treated as positives, at the risk of unnecessarily overloading the radiological review workflow.
Conclusion: Real Added Value, a Deployment That Requires Architecture
Available real-world data confirm that artificial intelligence applied to emergency radiography offers concrete, measurable, and clinically relevant added value, provided its strengths and limitations are understood with precision.
Its main strength lies in the detection of acute fractures, where its performance is comparable to that of a physician, with an exceptionally high negative predictive value that makes it a reliable diagnostic exclusion tool. Its ability to maintain this performance consistently across age groups and anatomical regions, including geriatric and pediatric populations, strengthens its credibility for routine deployment.
The detection of other traumatic injuries remains a point of attention for autonomous use. Certain anatomical variants generate recurring false positives, and pathologies under-represented in the training data are systematically under-detected. These points of vigilance are a reminder that deploying AI in emergency radiology must be accompanied by training users in its specific blind spots. The real question is no longer whether AI can be useful in emergency radiology. The data answer positively. The question now is: how to architect its integration to maximize clinical impact while preserving patient safety?
Toward a Mature Radiological AI: From Validation to Responsible Integration
The coming years will be decisive in transforming promising performance into demonstrated clinical benefit. Detection algorithms will tend to become increasingly capable.
For radiologists and emergency teams, the challenge is above all organizational and educational. Defining clear usage protocols, training teams in the critical interpretation of algorithmic outputs, and establishing feedback loops nourished by errors observed in real practice are the conditions for responsible deployment. The radiologist of the future will not simply be the one who uses AI: they will be the one who knows precisely when to trust it, and when to question it.
