Seventy FDA-cleared AI devices for orthopaedic surgery. Eight validated through a prospective clinical trial

That is the headline finding of a retrospective analysis published this month in JAAOS Global Research & Reviews by Lee, Jay, Fox, Padley, Dai, and Levin from Johns Hopkins Medicine. The study catalogues every FDA-cleared AI-enabled medical device with an orthopaedic-specific indication as of February 2025 - 70 devices in total - and systematically characterises their regulatory pathways, clinical validation, AI architecture, and commercialisation landscape. As a snapshot of where regulatory oversight stands for one of medicine's most procedurally intensive specialties, it is not a reassuring picture.

The growth numbers are striking. The three-year moving average of annual orthopaedic AIMD clearances increased more than fivefold from 2017-2019 to 2022-2024, rising from three devices per year to 16.6. Deep learning is now the dominant architecture, comprising 57% of devices cleared in the most recent period. Spine surgery accounts for 43% of cleared devices, followed by hip and knee at 20%, reflecting both the procedural complexity of these subspecialties and the existing infrastructure of imaging-based surgical planning tools on which AI augmentation is layered. Surgical planning is the predominant function across nearly all subspecialties, with the notable exception of trauma, where time constraints shift clinical need toward fracture identification and urgent surgical guidance.

None of this is surprising. What the study's regulatory analysis adds is a granular account of what "cleared" actually means in terms of the clinical evidence behind these devices, and the picture is considerably less robust than the word might imply.

More than 95% of AI medical devices across the FDA database are reviewed through the 510(k) pathway, which clears devices by demonstration of substantial equivalence to a predicate already on the market rather than by independent evidence of safety and efficacy. For orthopaedic AIMDs, 22.8% reached the market without any clinical testing at all - no human data, only bench testing of device software. An additional 68.6% were supported by retrospective clinical analyses, typically image repositories. Only 8.6% underwent a formal prospective clinical trial, and five of those six prospective studies were cleared between 2022 and 2024, with four appearing in 2024 alone - suggesting some recent movement in the right direction, but from a very low baseline. The clinical study populations are also highly variable: the mean sample size across all validation studies was 469 participants, but the median was 208, and the distribution was wide enough that the standard deviation exceeded the mean.

The improvement over time is genuine and worth acknowledging. The proportion of orthopaedic AIMDs cleared without any clinical data fell from 62% in 2017-2019 to 18% in 2022-2024. But improvement relative to a poor starting point is not the same as arriving at an adequate standard, and the paper is appropriately direct about this. Clearance summaries frequently omit the demographic characteristics of validation populations, randomisation or sampling methodology, and raw outcome measures - the variables that would allow any independent assessment of generalisability or bias risk. The authors flag this not as a peripheral concern but as structurally limiting: without mandatory public reporting of complete validation data, it is difficult to evaluate whether a cleared device is likely to perform equivalently across the diverse patient populations an orthopaedic practice actually serves.

One safety finding deserves to be read carefully rather than straightforwardly reassuring. No orthopaedic AIMD in this cohort has been subject to a recall or adverse event report. The authors correctly contextualise this: these devices have a short post-clearance history, and the recall-free record may reflect the relatively early stage of widespread deployment rather than confirmed long-term safety. It is also worth noting that almost all cleared orthopaedic AIMDs are software-based diagnostic and planning tools rather than devices that directly control physical interventions. The harm pathway for a flawed surgical planning recommendation is real but indirect, mediated by the surgeon's interpretation of the output. How that human-AI interaction performs under real-world clinical pressure, across different levels of AI literacy among orthopaedic surgeons, with patient populations that differ from the training and validation cohorts, remains essentially unstudied.

The commercialisation analysis adds another layer of context that regulators and healthcare systems should find informative. Private companies account for 61% of active manufacturers and develop their orthopaedic AIMDs almost entirely in-house (98% of private company devices). Publicly traded firms, by contrast, are substantially more likely to have acquired devices developed elsewhere, with established public companies developing only 53% of their portfolio internally and smaller public companies developing only 36% in-house. The implications of this pattern for quality control, post-market surveillance continuity, and regulatory accountability when devices change hands through acquisition have not been well studied, but the broader AIMD literature the authors cite suggests that publicly traded companies have historically driven substantially higher recall rates than privately held ones - a dynamic worth monitoring as the acquisition-driven commercialisation model grows.

The study is appropriately bounded in scope. It covers FDA-cleared devices only, not those under investigational review or cleared in other jurisdictions. It does not capture real-world surgeon adoption, clinical outcomes, or the growing category of AI tools embedded within larger platform systems rather than standing alone as discrete medical devices. These are not criticisms - they are honest acknowledgements of what this kind of regulatory database analysis can and cannot establish.

What it does establish, with reasonable clarity, is that the FDA clearance of an AI device for orthopaedic surgery is currently compatible with very limited prior clinical testing. The trajectory is improving. The standard it needs to reach - consistent prospective validation, transparent reporting of demographic composition, and systematic post-market surveillance - remains at some distance from where most of the market currently sits.

Full paper: https://pmc.ncbi.nlm.nih.gov/articles/PMC12879955/