Artificial Intelligence in Orthopaedic Surgery: An Evidence Review

Artificial intelligence in orthopaedic surgery is no longer a forecast. Software that marks a distal radius fracture on a wrist radiograph, segments a pelvic CT for hip planning or predicts the outcome of a joint replacement has been cleared for clinical use, and some of it now runs in NHS urgent care. The evidence behind each use is uneven.

This review covers what has been shown, in which patients, how the tools are regulated in the US and the UK, and where they fail. The short answer: fracture detection has the best evidence, as an aid to the clinician; outcome prediction mostly lacks external validation; and regulatory clearance is not proof of clinical benefit.

What "AI" means in these studies

Almost every tool in clinical orthopaedics is narrow. It does one defined task, on one kind of input, and was trained on labelled examples of that task.

Image models are mostly convolutional neural networks. They take a radiograph, CT or MRI and output a label ("fracture present"), a location (a box or heat map) or an outline of a structure (segmentation). Prediction models take tabular data (age, comorbidity, baseline scores) and output a probability, such as the chance of reaching a meaningful improvement in the Oxford Knee Score. Large language models generate text and are the newest arrival.

Two measures recur. Discrimination, usually the area under the receiver operating characteristic curve (AUC), says how well a model ranks patients with the outcome above those without. Calibration says whether a predicted 20% risk really means 20 in 100. A model can discriminate well and still be badly calibrated in a new hospital. Our guide to statistics for orthopaedic surgeons covers both.

Fracture detection on radiographs

Missed fractures in emergency care are the problem these tools were built for, and the first one to reach the US market was orthopaedic.

The founding reader study. Lindsey and colleagues trained a deep network on 135,409 radiographs annotated by 18 senior subspecialist orthopaedic surgeons. Emergency medicine clinicians then read wrist radiographs with and without it. Average sensitivity rose from 80.8% to 91.5% and specificity from 87.5% to 93.9%, a relative reduction in misinterpretation of 47%. The authors were affiliated with Imagen Technologies, which developed the software and funded the work.

The first FDA authorisation. On 24 May 2018 the FDA granted De Novo authorisation to Imagen OsteoDetect (opens in a new tab) for detecting distal radius fractures on posteroanterior and lateral radiographs of adult wrists. The submission rested on two retrospective studies: 1,000 images compared with three board-certified hand surgeons, and 24 clinicians reading 200 cases with and without the software. The FDA described it as an adjunct, not a replacement for the clinician's review.

That decision created a new device type, "radiological computer-assisted detection/diagnosis software for fracture" (21 CFR 892.2090, product code QBS). Its regulatory definition ends: "Primary diagnostic and patient management decisions are made by the clinical user."

Multi-region tools. Guermazi and colleagues tested BoneView on an external set of 480 examinations across eight body regions, with fracture prevalence fixed at 50% and 24 readers from six specialties. With AI help, sensitivity per patient rose from 64.8% to 75.2% and specificity from 90.6% to 95.6%, and reading time fell by 6.3 seconds per examination. The gain was not significant for the shoulder and clavicle or for the spine. Several authors were affiliated with Gleamer, the manufacturer.

AI against clinicians. The Kuo meta-analysis in Radiology (2022) pooled 42 studies, 37 of them on radiographs. On internal test sets, pooled sensitivity was 92% for AI and 91% for clinicians. On external test sets it was 91% for AI and 94% for clinicians, with specificity 91% and 94%. None of the differences was statistically significant, and 22 of the 42 studies (52%) had a high risk of bias.

Fracture-detection devices on the FDA's AI-enabled device list include:

Source: FDA AI-Enabled Medical Devices list and 510(k)/De Novo databases. BoneView and Rayvolve have later clearances for updated versions.

The UK position. In January 2025 NICE published early value assessment guidance (HTG739, formerly HTE20) (opens in a new tab) on AI to help detect fractures on X-rays in urgent care. Four technologies can be used in the NHS during a two-year evidence-generation period: Rayvolve and TechCare Alert for people of any age, and BoneView and RBfracture for people aged 2 and over. Each also needs appropriate regulatory approval, including NHS England's Digital Technology Assessment Criteria (DTAC). A fifth product, qMSK, is restricted to research. The committee noted that the evidence came from retrospective studies, with no data on the costs and clinical outcomes of missed fractures.

Predicting outcomes after joint replacement

The second large body of work uses registry and health-record data to predict how a patient will do after hip or knee arthroplasty.

Patient-reported outcomes. Fontana and colleagues used the Hospital for Special Surgery registry (7,239 hip and 6,480 knee replacements) to predict who would fail to reach the minimal clinically important difference in four PROMs at two years. Three algorithms (LASSO logistic regression, random forest and a support vector machine) performed almost identically, with AUCs of 0.60 to 0.89. The baseline score was consistently among the strongest predictors. The authors warned that the models "cannot simply be applied off-the-shelf without proper testing".

Huber and colleagues trained eight classifiers on 130,945 NHS PROMs records (April 2015 to April 2017). The best AUCs were about 0.78 for improvement in the Oxford Hip Score and 0.70 for the Oxford Knee Score. The most important predictors were again the preoperative scores. Gradient boosting performed best, but a plain linear model was among the top performers. Our guide to PROMs and the MCID explains why the threshold you choose changes these results.

Length of stay. Ramkumar and colleagues applied a naive Bayesian model to 122,334 primary total hip replacements from a US administrative database. It predicted length of stay (grouped as 1-2, 3-5 or 6+ days) with an AUC of 0.87, and inpatient payment with an AUC of 0.71. Payment errors grew with comorbidity: 3%, 12% and 32% for moderate, major and extreme comorbidity.

Three lessons follow. Most of the predictive signal comes from variables a surgeon already records, above all the baseline score. Machine learning often adds little over regression on this kind of data. And a model built on one US hospital system or one national dataset says little about your own patients until someone tests it there. Our registries and big data piece covers what these datasets can and cannot show.

Segmentation and pre-operative planning

Segmentation is outlining a structure in an image so it can be measured or turned into a 3D model. Done by hand on CT it is slow, and it is the step that 3D planning, patient-specific instrumentation and 3D-printed guides and implants depend on.

Knee MRI. Norman and colleagues trained a U-Net on 638 knee MRI volumes. Overlap with manual segmentation (Dice coefficient, where 1.0 is perfect) was 0.770 to 0.878 for cartilage compartments, 0.809 for the lateral meniscus and 0.753 for the medial meniscus. Each segmentation took about 5 seconds, and automated volume and thickness measurements correlated with manual values at 0.93 to 0.94.

Hip CT planning. Formus Hip, CT-based planning software for hip arthroplasty, was cleared by the FDA in March 2023 (K213272) (opens in a new tab). Its 510(k) summary reports:

• an AI segmentation algorithm validated on CT scans from 60 US patients against manual segmentation by radiologists, with mean Dice scores of 0.95 for the hemipelvis and 0.97 for the femur
• implant sizing tested on 133 CT scans from Australia against surgeons' templating, with the acceptance target set at 80% of sizes within two sizes; more than 90% of recommended stem and cup sizes met it
• the statement that "substantial equivalence was not based on an assessment of clinical performance data"

That last line is the norm, not the exception. A 510(k) clearance shows that the software does what it claims, technically, as well as a predicate device. It does not show better implant position, fewer dislocations or better PROMs. Within two sizes is also a wide margin for a cup. The same gap between technical accuracy and clinical benefit runs through robotic arthroplasty and surgical navigation.

Language models and ambient scribes

Knowledge tests. Lum gave ChatGPT 207 text-only questions from the Orthopaedic In-Training Examination in 2023. It answered 47% correctly (opens in a new tab), around the 40th percentile for first-year residents and the first percentile for final-year residents, and its accuracy fell from 54% to 34% as questions moved from recall to application. Newer models will score differently. The finding to keep is that performance falls as questions demand more judgement.

Ambient scribes. These tools record the consultation and draft the clinic letter. In a pragmatic randomised trial published in NEJM AI in 2025, 238 outpatient physicians from 14 specialties were allocated to one of two scribes or usual care. One scribe (Nabla) reduced time spent writing notes by 9.5%; the other (DAX Copilot) produced no significant change. Clinicians reported clinically significant inaccuracies "occasionally" with both. Claims that scribes save hours a day are not supported by that trial.

In England, NHS England's guidance on AI-enabled ambient scribing products (opens in a new tab) expects a clinical safety case (DCB0160), a data protection impact assessment and clinician review of every output. It warns that accuracy may vary with accents, dialects and speech impairment, and that NHS organisations may remain liable where a non-delegable duty of care is involved. Using unapproved public chatbots with patient information falls outside that guidance. For research use, see what journals now require on AI-assisted writing.

How AI devices are regulated

United States. The FDA regulates AI software that meets the definition of a medical device. Most orthopaedic imaging tools reach the market as class II devices through the 510(k) route (substantial equivalence to an existing device) or, for a new type, De Novo. The FDA publishes an AI-Enabled Medical Devices list (opens in a new tab). In the version downloaded for this review (page updated September 2026), it held 1,614 authorisations, of which 1,230 had radiology as the lead review panel and 8 had orthopaedics. The FDA states that the list is not comprehensive, because devices are identified by AI-related terms in their summaries. In December 2024 the FDA issued final guidance on predetermined change control plans, which let a manufacturer specify in advance how a cleared AI model may be updated.

United Kingdom. The MHRA regulates software, including AI, as a medical device when it has a medical purpose. Its Software and AI as a Medical Device Change Programme (opens in a new tab) roadmap, published on 17 October 2022, covers qualification, classification, transparency and retraining of AI models. In May 2024 the MHRA launched the AI Airlock (opens in a new tab), a regulatory sandbox for AI medical devices that initially supported four to six projects. Market approval and NHS adoption are separate steps: NICE assesses value, and products used in the NHS must also meet NHS England's Digital Technology Assessment Criteria (DTAC).

Why a good AUC can fail in your hospital

Dataset shift. A model learns its training data, including features that have nothing to do with the disease. Zech and colleagues trained pneumonia detectors on 158,323 chest radiographs from three US hospital systems. The networks could identify which hospital a radiograph came from in 99.95% and 99.98% of cases. A model with an internal AUC of 0.931 fell to 0.815 at an external hospital.

Shortcut learning in hip fractures. Badgeley and colleagues trained models on 17,587 pelvic radiographs. The image alone predicted hip fracture with an AUC of 0.78. Adding patient and hospital-process data raised it to 0.91, and the radiograph by itself predicted the scanner model with an AUC of 1.00. On a test set balanced for patient and process variables, fracture prediction fell to 0.52, no better than chance. The model had learned who gets a pelvic radiograph and how, not what a fracture looks like.

Little external validation. Groot and colleagues found that only 10 of 59 machine-learning models predicting orthopaedic surgical outcomes had been externally validated, in 18 studies. All retained good discrimination, but only 3 studies reported other key measures such as calibration, and median completeness against the TRIPOD reporting checklist was 61%.

Bias. Image classifiers trained on large chest radiograph datasets underdiagnosed under-served groups (opens in a new tab) such as female, Black and low-income patients, and underdiagnosis was worse still in intersectional subgroups. The same risk applies to any musculoskeletal model trained on a narrow population, which is why NICE lists accuracy in different populations among its evidence gaps.

Reporting standards. TRIPOD+AI (2024) replaces the 2015 TRIPOD checklist with 27 items for any prediction model study, whether it uses regression or machine learning. Use it as a checklist when you read one.

Questions to ask before you rely on a tool

1. What exactly is it cleared for? OsteoDetect covers distal radius fractures in adults. NICE limits two fracture tools to patients aged 2 and over. Outside the intended use, you are on your own.
2. Was it validated on patients like yours? Look for external, ideally prospective, testing at sites with different scanners and case mix.
3. Standalone or reader study? What matters is whether clinicians using the tool do better, not whether the tool beats a reference standard on a curated set.
4. Discrimination and calibration? An AUC alone does not tell you whether a predicted risk can be quoted to a patient.
5. Who ran the study? Several pivotal studies were done by the manufacturers. That does not make them wrong, but it raises the bar for independent replication.
6. What happens locally when it is wrong? Someone must audit performance after go-live, as the NHS England scribe guidance requires.

For general appraisal skills, see how to read a research paper critically. For what all this means for surgical careers, see our opinion piece, Will AI Replace Surgeons?