China has approved 154 AI medical devices since 2020, and 69% of them read scans
A new audit of China's regulatory record finds a market concentrated in radiology, dominated by deep learning, and clustered in four cities. It also finds the approval curve flattening.
Most of what gets written about medical AI regulation describes the US Food and Drug Administration's clearance list. A retrospective analysis published on 18 March does the same exercise for China, and the shape of the market it describes is different in ways that matter [s1].
The authors searched the Drugdataexpy database for artificial-intelligence-based medical devices approved in China through 30 June 2025, using AI-related keywords in the "structural composition" and "intended use" fields, then manually verified each hit and excluded non-AI devices [s1]. What survived: 154 devices, approved from the first in 2020 onward [s1].
The curve
Annual approvals rose from 9 in 2020 to 45 in 2024 — a compound annual growth rate of 49.53% [s1].
Then the flattening. The first half of 2025 produced 20 approvals [s1]. Doubled, that is a 2025 pace roughly level with 2024 rather than continuing the climb, which the authors read as a potential moderation [s1]. It is half a year of data and should be treated as such; a single half-year does not establish a trend reversal. But it is the first interruption in five years of acceleration.
What kind of devices
Two concentrations dominate.
The first is clinical. Radiology accounts for 106 of the 154 approvals, or 68.8% [s1]. Computed tomography is the primary data source for 96 devices, 62.3% of the total, particularly for pulmonary nodule detection and cardiovascular assessment [s1]. Nearly seven in ten approved Chinese AI medical devices look at pictures, and most of those pictures are CT scans.
The second is technical. Deep learning is the algorithm type in 143 of 154 devices — 92.9% [s1]. There is very little methodological diversity in what has cleared.
Neither concentration is unique to China. Both are worth stating precisely, because "AI in medicine" tends to be discussed as if it spanned the whole of clinical practice, and the approved reality in the world's second-largest device market is overwhelmingly one specialty reading one modality.
How they got through
China classifies devices by risk, and the AI cohort sits high: 123 of 154 devices, 79.9%, are class III, the highest-risk category [s1]. Risk class was significantly associated with approval year (P=.03), manufacturer location (P=.03) and medical specialty (P=.004) [s1].
The evaluation pathway tracks the class. Clinical trials were the primary evaluation route for 118 of 154 devices overall (76.6%), and for 116 of the 123 class III devices (94.3%) [s1]. Most class II devices went the other way: 21 of 31, 67.7%, used a clinical exemption pathway [s1].
Expedited routes were used sparingly. Of the 123 class III devices, 19 (15.4%) were approved through innovation review, and 2 each (1.6%) through priority and emergency approval [s1].
That combination — a mostly class III cohort, cleared mostly on clinical trial evidence, with expedited pathways used for a minority — describes a regime that the authors characterise as risk-proportionate [s1]. It is a meaningfully different evidentiary posture from a market where most AI devices clear on comparison to an existing product.
Who is building them
Market concentration is the study's other headline finding. The top four manufacturers account for 59 of 154 approvals, 38.3% of the total [s1]. Geographically, developers cluster in Beijing, Shanghai, Shenzhen and Hangzhou [s1].
Limits worth holding onto
This is a database study of regulatory records, not of clinical performance. It counts approvals; it does not measure whether any of these 154 devices improved a patient outcome, how many are actually deployed in Chinese hospitals, or how they perform outside the populations they were validated in. The authors frame their contribution as an up-to-date overview of approval trends and characteristics, and that is what it supports [s1].
The dataset also stops at 30 June 2025, and derives from a commercial database rather than a regulator's own published register, with manual verification as the quality control step [s1]. Keyword-based identification of AI devices will miss products whose registration text does not describe the algorithm, and may over-capture others — which is why the manual exclusion step exists, and why the count should be read as a careful estimate rather than an exact register.
What to watch
Whether the 2025 moderation holds through full-year figures, and whether the radiology-and-CT concentration starts to loosen as approvals accumulate in other specialties. Both are answerable from the same public record this study used, which is the useful thing about counting approvals: the next year's data arrives on its own.
This article is informational and does not constitute medical advice.
Sources
- [s1] Zhang L, Yan J., "Approval of AI-Based Medical Devices in China From 2020 to 2025: Retrospective Analysis," JMIR Medical Informatics, published online 18 March 2026. https://doi.org/10.2196/85538
Sources
- Approval of AI-Based Medical Devices in China From 2020 to 2025: Retrospective Analysis — JMIR Medical Informatics , March 18, 2026
FDA opens comment on how to regulate AI medical devices that generate their own answers
A new discussion paper proposes a two-axis risk framework and a physician-training analogy for evaluating generative AI devices. It is not a rule, and the agency is asking what one should look like.
The FDA has authorised 723 AI radiology tools. 33 were tested prospectively.
A systematic review of every AI device authorised from 1995 to June 2024 found 97% cleared through the 510(k) pathway, which does not require independent clinical data on performance or safety.
95.5% of FDA AI device summaries omit who the model was trained on
A census of 691 AI-enabled devices cleared through 2023 found demographic data missing almost everywhere, three summaries reporting patient outcomes, and 113 recalls driven mostly by software.
The FDA has cleared 1,357 AI medical devices. Three were tested on patient outcomes.
A researcher who expected the evidence base to be thin says even she was surprised by how thin. Most cleared devices never appear in a registered clinical trial at all.