FDA clearance for AI medical devices from Korea, Japan and Taiwan almost always comes through a 510(k): 173 of the 174 such devices on FDA's AI list did. Your MFDS, PMDA or TFDA dossier gives you a strong start, but FDA adds its own expectations, above all validation data that represents US patients and US medical practice.
Key takeaways
- FDA's list of AI-enabled devices includes 81 from Korean applicants, 63 from Japanese and 30 from Taiwanese. Korea ranks 5th of all countries and Japan 6th.
- FDA's draft AI guidance says at least three geographically diverse US clinical sites may be appropriate for clinical validation. Of the 2025 Korean, Japanese and Taiwanese AI clearances that said where their data came from, 17 of 18 used US data.
- For AI triage, detection and diagnosis software, FDA's special controls spell out the testing: standalone performance, enough cases in each important subgroup and, for detection and diagnosis aids, evidence that readers improve.
- All three home regulators have change plans for AI updates, yet none of the 29 Korean, Japanese and Taiwanese AI devices cleared in 2025 came with an FDA PCCP. 31 of all 329 AI devices cleared that year did.
- AI 510(k)s took a median of 171 days for Korean applicants, 130.5 for Japanese and 208 for Taiwanese, against 154 for all AI devices.
How do AI devices from Korea, Japan and Taiwan reach the US market?
Almost always through a 510(k). FDA's public list of AI-enabled devices held 1,614 devices in the version current as of September 4, 2026 [1]. We matched each device to its applicant's country in FDA's data [1]:
| Applicant country | AI devices on FDA's list | Rank by country | Most common device types (product code) |
|---|---|---|---|
| South Korea | 81 | 5 | Image processing software (QIH, LLZ), ultrasound (IYN), triage (QAS) |
| Japan | 63 | 6 | CT (JAK), MRI (LNH), ultrasound (IYN) |
| Taiwan | 30 | 11 | Triage (QAS), radiotherapy software (QKB) |
Only the US, Israel, France and China had more than Korea [1]. Korea's AI devices are mostly standalone software. Japan's are mostly AI features inside CT and MRI scanners [1].
Of the 174 devices from these three countries, 173 were 510(k) clearances and one was a De Novo grant, from Korea [1]. So the first job is the same as for any 510(k): find your product code and a predicate legally marketed in the US. Your MFDS, PMDA or TFDA approval does not replace either.
The product code also tells you which regulation applies, and some AI regulations come with detailed special controls [6]:
| Regulation | Device type | Common product code | AI-specific special controls? |
|---|---|---|---|
| 21 CFR 892.2050 | Medical image management and processing system | QIH, LLZ | No. Its special controls are imaging standards such as DICOM |
| 21 CFR 892.2060 | Computer-assisted diagnosis software for lesions suspicious of cancer | Varies | Yes |
| 21 CFR 892.2070 | Medical image analyzer, including computer-aided detection | Varies | Yes |
| 21 CFR 892.2080 | Radiological computer-aided triage and notification software | QAS | Yes |
Complizen's Device Classification tool returns your device's product code and regulation, with the special controls and FDA guidance that apply.
What carries over from your MFDS, PMDA or TFDA dossier?
A lot of the evidence, and the logic behind your change plan. FDA does not accept a home approval in place of a 510(k), but the regulators ask many of the same questions:
| What FDA looks for | What your home regulator already asked | What to add for FDA |
|---|---|---|
| A description of the model, its inputs, outputs and limits [3, 6] | TFDA's AI guidance asks whether the algorithm is locked or adaptive and how it was trained [11]. MFDS labeling must describe the training method and data [9] | Little, if your description is complete. Translate it and align it with FDA's terms |
| Test data independent of training data [3] | TFDA requires test data kept strictly separate from training and validation data [11] | Show how the test data were sequestered from the model developers [3] |
| Validation that represents the intended use population [3] | TFDA asks why non-Taiwan test data apply to Taiwan [11]. MFDS asks AI products for performance data based on real health data [9] | US data, or a written explanation of why your data represent US patients and US practice [3] |
| Subgroup results [3, 6] | Varies by regulator | Results stratified by important covariates, such as age, sex, race, ethnicity and imaging equipment [3, 6] |
| A plan for model updates [4] | MFDS change management plan [9], Japan's IDATEN change plan [10], TFDA's PCCP guidance [11] | An FDA PCCP with its own three sections [4] |
| A US predicate [12] | Not asked | A cleared device with the same intended use |
Our guides to what an MFDS licence changes, what a PMDA approval changes and what a TFDA licence changes cover the non-AI side of each.
Does FDA accept validation data from Korean, Japanese or Taiwanese hospitals?
It can, but you have to explain it, and you may need US data too. This is the biggest gap between an Asian AI dossier and an FDA submission.
FDA's draft guidance on AI-enabled device software, published in January 2025, gives the clearest signal [3]:
- Data collected outside the US can bias a model "if the OUS population does not reflect the U.S. population due to differences in demographics, practice of medicine, or standard of care."
- "The use of at least three geographically diverse US clinical sites (or health care systems) may be appropriate to clinically validate an AI-enabled device."
- If you use non-US data in validation, explain how they compare with the US population and US medical practice, including disease presentation, prevalence and progression.
- AI devices "may require higher proportion of U.S. data in the clinical validation," and FDA encourages a Q-Submission to get feedback on your planned use of non-US data.
The guidance is still a draft, but cleared devices show what applicants do in practice. We read all 29 510(k) summaries for AI devices from Korean, Japanese and Taiwanese applicants cleared in 2025 [2]:
| What the summary said about data origin | Summaries |
|---|---|
| Included US data, either alone or with data from the home country or elsewhere | 17 |
| Used home-country data only (one Korean device) | 1 |
| Did not say where the data came from | 11 |
The patterns are practical. One Japanese MRI reconstruction device was tested on subjects from sites in the US and Japan, and its summary states why the population represents the intended US population [2]. One Korean triage device was trained on data from the US, Brazil and Korea, then tested on 560 scans from US regions [2]. One Taiwanese contouring tool was trained on Taiwanese data and tested on cases from three US institutions and one non-US institution [2].
Subgroups matter too. Of all 329 AI-device 510(k) summaries from 2025, 141 (43%) mentioned subgroup analysis and 105 (32%) mentioned race or ethnicity [2].
This is not only an Asian question. More than half of the AI devices on FDA's list with a known applicant country came from outside the US: 855 of 1,593 [1]. Israel (125), France (96) and China (85) each have more than Korea [1]. In 2025, 194 of the 329 AI devices cleared through a 510(k) came from applicants outside the US [2]. Every one of them faces the same question about US data.
If your validation data are mostly from outside the US, the safest route is to ask FDA early. A Pre-Submission lets you put your data plan in front of FDA before you spend on testing [3, 8]. Our Q-Submission guide explains how.
What testing does FDA require for AI radiology software?
For three common AI device types, the regulation itself lists the testing [6]:
- Triage and notification software (21 CFR 892.2080). You need a detailed description of the algorithms, pre-specified testing that shows effective triage, and standalone performance results. The test dataset must have enough cases from important cohorts to estimate performance for each subset [6].
- Detection aids (21 CFR 892.2070). Testing must show that the device improves reader performance in the intended use population. It can use standalone tests, side-by-side comparisons or a reader study, as applicable [6].
- Diagnosis aids for lesions suspicious of cancer (21 CFR 892.2060). Testing that shows the device improves reader performance, plus standalone performance results [6].
The labeling requirements are just as specific. For triage software, labeling must describe the patient population, the intended user and their training, and situations where the device may fail. It must also cover compatible imaging hardware and summarize performance testing, with subgroup results [6].
Software under 21 CFR 892.2050, which covers most image processing software, has no AI-specific special controls [6]. FDA's draft AI guidance still applies to its performance evidence [3]. Among 2025 AI-device summaries, 40 of 329 (12%) mentioned a reader study [2].
How do you handle AI model updates at FDA?
Through a predetermined change control plan (PCCP), or a new 510(k) for each significant change.
Section 515C of the FD&C Act lets FDA authorize a PCCP as part of your 510(k). Changes made under an authorized PCCP do not need a new submission [4]. FDA's final PCCP guidance for AI-enabled devices asks for three sections [4]:
- Description of Modifications. The planned changes and their specifications.
- Modification Protocol. The verification and validation methods and acceptance criteria for those changes.
- Impact Assessment. The benefits and risks of the planned changes and how you will reduce the risks.
PCCPs are becoming part of AI clearances. In 2025, 31 of the 329 AI devices cleared through a 510(k) had a PCCP in their clearance [2]. None of the 29 from Korea, Japan or Taiwan did [2]. That gap stands out, because these manufacturers already run change plans at home. Korea has MFDS's change management plan [9]. Japan has had IDATEN since September 1, 2020 [10]. Taiwan has TFDA's PCCP guidance of September 2024 [11]. Each covers similar ground. Your home plan is a good draft, but it needs rewriting into FDA's three sections and FDA's acceptance criteria.
Without a PCCP, you judge each change against FDA's 2017 guidance on software changes [5]. If a change needs a new 510(k), a Special 510(k) is often the faster route for your own cleared device. From 2023 to June 2026, AI Special 510(k)s cleared in a median of 29 days [1]. Our guide to whether a PCCP is only for AI devices covers the scope.
What else does FDA add for AI software?
Three areas often catch Asian AI manufacturers by surprise:
- Cybersecurity. If your software can connect to the internet, it may be a "cyber device" under section 524B of the FD&C Act. Your 510(k) must then include the cybersecurity information that section requires, including a software bill of materials covering commercial, open-source and off-the-shelf components [7]. FDA updated its cybersecurity guidance on February 3, 2026 [7].
- Software documentation level. Your 510(k) states a documentation level, Basic or Enhanced, and justifies it. Our guide to FDA's software documentation level explains how to decide.
- US labeling. Where special controls apply, they set what your labeling must say about population, users, limits and performance [6]. Your home labeling is a starting point, not a substitute.
How long does FDA take to clear AI devices from these countries?
We measured AI 510(k)s cleared from January 2023 to June 2026, from FDA's receipt to clearance [1]:
| Applicant country | Traditional AI 510(k)s | Median days |
|---|---|---|
| Japan | 32 | 130.5 |
| United States | 347 | 147 |
| All countries | 823 | 154 |
| South Korea | 52 | 171 |
| Taiwan | 17 | 208 |
Device type explains part of the spread. For automated image processing software (QIH), Korean, Japanese and Taiwanese applicants took a median of 183 days, against 153.5 for all AI applicants. For triage software (QAS), they took 155 days against 110 [1]. Most of a long total is usually time on hold while FDA waits for answers, so a complete data story up front saves the most time. Our review times by country analysis covers all device types.
What should you do next?
- Confirm your product code and regulation. They decide your predicate options and whether special controls set your testing.
- Map your home dossier against FDA's expectations. The table above is a starting point: model description, data independence, subgroups, US representativeness, labeling and cybersecurity.
- Plan your US data early. If most of your validation data come from Korea, Japan or Taiwan, ask FDA through a Pre-Submission before you test.
- Turn your home change plan into an FDA PCCP. Submit it with your first 510(k), so routine model updates do not need new submissions.
If you want a second view of step 2, Complizen's free Gap Assessment checks your device documents against what a 510(k) needs before you file.
Frequently asked questions
How do AI medical devices from Korea, Japan and Taiwan get FDA clearance?
Almost always through a 510(k). Of 174 AI devices from these three countries on FDA's AI list, 173 were 510(k) clearances and one was a De Novo grant. You need a US predicate, performance data that represent US patients and practice, and, for many device types, testing set by FDA's special controls.
Does FDA accept AI validation data from Korean or Japanese hospitals?
It can, if you explain how the data compare with US patients and US medical practice. FDA's draft AI guidance says AI devices may need a higher share of US data, and at least three geographically diverse US sites may be appropriate. Most Asian AI devices cleared in 2025 included some US data.
How much US data does FDA want for an AI device?
FDA sets no fixed share. Its draft AI guidance says at least three geographically diverse US clinical sites may be appropriate for clinical validation, and that AI devices may need a higher proportion of US data. FDA encourages a Pre-Submission to agree your planned use of non-US data before testing.
Does an MFDS, PMDA or TFDA approval help with FDA clearance?
Not directly. FDA does not accept these approvals in place of a 510(k). But the evidence behind them, such as your model description, separate test data and performance results, often carries over. It usually needs adding to, especially for US representativeness, subgroup results and cybersecurity.
Can my home-country change plan become an FDA PCCP?
It can be the starting point. MFDS, PMDA and TFDA change plans cover similar ground, but FDA's PCCP has three set sections: description of modifications, modification protocol and impact assessment. None of the 29 Korean, Japanese and Taiwanese AI devices cleared in 2025 had an FDA PCCP.
How long does FDA take to clear AI devices from these countries?
From 2023 to June 2026, Traditional AI 510(k)s took a median of 130.5 days for Japanese applicants, 171 for Korean and 208 for Taiwanese, from receipt to clearance. The median for all AI devices was 154 days. Much of a long total is usually time on hold.
What testing does FDA require for AI triage software?
Triage software under 21 CFR 892.2080 must include a detailed algorithm description, pre-specified testing showing effective triage, standalone performance results and a test dataset large enough to estimate performance in important subgroups. Its labeling must summarize performance testing, including subgroup results.
Do AI detection aids need a reader study?
Often. Under 21 CFR 892.2070, testing must show the device improves reader performance in the intended use population, using standalone tests, side-by-side comparisons or a reader study as applicable. In 2025, 12% of AI-device 510(k) summaries mentioned a reader study.
Is my AI software a cyber device?
It may be. Under section 524B of the FD&C Act, a cyber device includes software, can connect to the internet and could be vulnerable to cybersecurity threats. Its 510(k) must include the cybersecurity information section 524B requires, including a software bill of materials.
How do Korea, Japan and Taiwan regulate AI devices compared with FDA?
All four regulate AI software as a medical device by risk, and all four have a way to pre-approve model updates. Korea uses its Digital Medical Products Act, Japan the PMD Act with the IDATEN change plan, and Taiwan its Medical Devices Act with a PCCP guidance. FDA uses the PCCP under section 515C.
Should I hold a Pre-Submission meeting for an AI device?
Often, yes. FDA's draft AI guidance encourages a Q-Submission to get feedback on non-US data, and FDA's PCCP guidance encourages one before you submit a PCCP. Both are common questions for Asian AI manufacturers. An early answer from FDA can save a whole round of testing, so plan the meeting before you lock your data.
How many AI devices from Korea, Japan and Taiwan has FDA cleared?
FDA's AI-enabled device list includes 81 devices from Korean applicants, 63 from Japanese and 30 from Taiwanese. Korea ranks 5th of all countries, Japan 6th and Taiwan 11th. Radiology accounts for most of them, and all but one went through a 510(k) rather than De Novo or PMA.
References
- FDA. Artificial Intelligence-Enabled Medical Devices, list current as of September 4, 2026, matched to applicant country and review times through openFDA's 510(k) and PMA APIs. Accessed October 13, 2026.
- FDA. 510(k) Premarket Notification database: 510(k) summaries and clearance letters for the 329 AI-enabled devices cleared in 2025, read by Complizen. Accessed October 13, 2026.
- FDA. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. Draft guidance, January 2025.
- FDA. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions. Final guidance, issued December 4, 2024, updated August 18, 2025.
- FDA. Deciding When to Submit a 510(k) for a Software Change to an Existing Device. Final guidance, October 2017.
- eCFR. 21 CFR 892.2050, 892.2060, 892.2070 and 892.2080. Current as of October 2026.
- FDA. Cybersecurity in Medical Devices: Quality Management System Considerations and Content of Premarket Submissions. Final guidance, February 3, 2026.
- FDA. Requests for Feedback and Meetings for Medical Device Submissions: The Q-Submission Program. Final guidance, May 29, 2025.
- Korea: Digital Medical Products Act (디지털의료제품법), in force January 24, 2025, with its Enforcement Rule; MFDS press releases on the change management plan guideline, June 26, 2026 and the approval and review rules, July 2026.
- Japan: PMD Act; MHLW notice on the change plan system, August 31, 2020; MHLW Q&A on change plans and AI devices, December 22, 2023.
- Taiwan: TFDA AI/ML device software registration guidance, revised August 17, 2021; AI/ML CADe and CADx guidance, revised August 2025; PCCP guidance, September 18, 2024.
- eCFR. 21 CFR 807.92, content of a 510(k) summary, including the legally marketed predicate. Current as of October 2026.