AI Healthcare Regulations in Japan: Medical Device Approval

Introduction: 

When an algorithm can read a chest scan, flag a polyp on an endoscopy feed, or suggest a diagnosis, it stops being just software. In Japan, it becomes a regulated product with a legal path to the clinic. AI healthcare regulations in Japan are built on that idea, and understanding it is the first step for any company hoping to sell medical AI in the country.

A framework built on existing law

Japan did not write a separate AI law for medicine. It adapted the Pharmaceuticals and Medical Devices Act, known as the PMD Act. A 2013 amendment brought programs, meaning software that is not a physical object, into the legal category of medical devices. Since that change, which took effect in November 2014, software intended to diagnose or treat disease has been regulated as software as a medical device, or SaMD.

The division of labor is simple. The Ministry of Health, Labour and Welfare (MHLW) sets policy and grants final marketing approval. The Pharmaceuticals and Medical Devices Agency (PMDA) does the technical work of reviewing the evidence and advising developers. Anyone studying Japan software as a medical device regulations will find these two bodies at the center of almost every process.

Not every health app is swept in. Software intended for diagnosis, treatment, or prevention is regulated as a medical device when it falls into Class II or higher, while low-risk Class I software sits outside the Act. Wellness coaching, educational programs, and hospital administration tools generally fall on the unregulated side. Because the boundary can be hard to judge from the outside, the PMDA offers a consultation specifically to confirm whether a given product qualifies as a medical device at all.

Risk decides the route

Japan sorts medical devices into four risk classes, and the class determines who looks at the product before it reaches patients. Lower-risk devices can be self-declared or certified by a third party. Higher-risk devices need approval from the minister after PMDA review.

For artificial intelligence, the practical rule is more demanding than the class chart suggests. Where the Minister has set certification standards for a device category, certification can replace approval. Early machine-learning software fit no such category, and it was treated as unprecedented, so it required full approval. Where certification does not apply, approval by the PMDA is required. A company bringing a novel AI diagnostic tool to Japan should therefore plan for PMDA review rather than hoping for a lighter route.

That review has a track record behind it. The first AI-based medical device in Japan, software that helps distinguish tumor from non-tumor colorectal lesions on endoscopic images, was approved in December 2018. By September 2025, 51 of the 172 approved program medical devices carried an AI designation, spanning imaging, endoscopy, cardiology, and infectious disease. Japan regulates AI as an established category, not an experiment.

The road to PMDA approval

The PMDA approval process for AI diagnostic software begins before any application is filed. Developers are encouraged to settle two questions early: whether the product is a device, and how it should be classified. From there, the PMDA’s consultation services help teams test their study designs and evidence plans against what reviewers will expect. A dedicated SaMD consultation category, created under the government’s DASH for SaMD 2 strategy, exists precisely because digital products do not behave like pills or scalpels.

Once a dossier is submitted, reviewers examine clinical utility, performance, and safety. That dossier typically covers the product’s intended use, how the algorithm was built, the data used to train and test it, validation results, risk management, and quality management. A positive review leads to approval by the minister, and only then can the product be marketed.

There is one more hurdle that many founders underestimate. Adoption of AI-enabled devices depends significantly on inclusion in the National Health Insurance benefit schedule. A product that has regulatory clearance but no reimbursement route may find few hospitals willing to buy it. The most successful teams treat approval and reimbursement as one strategy, not two.

What regulators want to see

Japan regulatory requirements for AI medical devices tend to circle around a small number of recurring concerns. The first is intended use. A tightly defined purpose, such as supporting radiologists in detecting a specific finding on a specific scan type, gives reviewers something concrete to evaluate. Vague ambitions invite vague evidence and slow reviews.

The second concern is data. Regulators look at where the data came from, how it was labeled, how training and test sets were separated, and whether the population reflects the patients who will actually be treated. Developers also have to work within a wider data environment. In September 2024, the MHLW issued guidelines on using medical digital data for AI research and development, and the country’s privacy rules and the Next Generation Medical Infrastructure Act shape how health data can be handled.

The third concern is real-world performance. Clinical evidence for AI-enabled medical software in Japan is strongest when it shows the tool working in the setting where it will be used, ideally with Japanese clinical sites or patients, and with results broken down across subgroups such as age, sex, equipment type, and hospital. A high accuracy figure on a tidy test set is a beginning, not a conclusion.

That evidence also has to be communicated. A 2025 review of PMDA-approved AI radiology products concluded that current package insert requirements offer insufficient transparency for evidence-based clinical implementation. The finding is a reminder that regulators and clinicians are both asking for clearer disclosure, and companies that publish detailed performance data are likely to find the market more receptive.

Approval in two steps

One of the more distinctive features of Japan’s approach is its two-step approval concept for SaMD. The idea is that a software product can receive a first-stage approval for a limited scope, then a second-stage approval is expected once post-marketing clinical studies and real-world data have accumulated.

The logic suits software. A digital tool can reach patients earlier under a carefully bounded claim, while evidence grows in real use. The government organized and publicized the approach in DASH for SaMD 2, a strategy announced on 6 September 2023 for the following five years. That strategy also expanded the PMDA’s SaMD review office into a full department, and it set out to support the overseas expansion of Japan-originated software. The PMDA has also trialed a priority review designation for innovative SaMD, especially tools that use advanced technology such as artificial intelligence.

Letting AI change after approval

The deepest regulatory puzzle with AI is that it is meant to improve. A traditional device is approved in a fixed form, and any significant modification means a new application. For software that learns, or simply gets retrained as new data arrives, that arrangement can freeze progress.

Japan’s answer is IDATEN, short for Improvement Design within Approval for Timely Evaluation and Notice. It came out of the PMD Act amendment published on 4 December 2019. Under IDATEN, a manufacturer submits a plan for expected modifications in advance, and the regulator approves the plan itself. When a change is needed later, it can be carried out by notification instead of a full partial-change application, which saves considerable time.

The approach shifts regulatory attention from each individual update to the discipline behind all of them. To use it, a developer must state what will change, how every change will be validated, and what limits will keep performance inside approved bounds. Done properly, it lets a product evolve without sacrificing the safety review that justified approval in the first place. Japan’s scheme is often compared with the predetermined change control plans discussed elsewhere, and it has been an influential early example.

The generative AI question

If the framework handles conventional AI with growing confidence, generative AI is where certainty runs out. Large language models and similar systems do not produce fixed outputs. They can vary from one run to the next, and they can state errors fluently. Those traits complicate nearly every classical step of device review, from validation to risk control.

Regulatory guidance for generative AI medical software in Japan is still taking shape. A 2026 policy paper by authors associated with the MHLW and PMDA notes that deliberations on generative AI have only recently begun within a research project commissioned by the ministry. In other words, the rules are being written while the technology is already entering clinical conversations.

For developers, the safest working assumption is that a generative tool with a diagnostic, therapeutic, or preventive purpose can qualify as a medical device, and that early PMDA consultation is the sensible way to find out. Responsibility is another open matter. One recent review observes that Japan’s framework sets pre-market performance requirements but does not yet specify post-market liability norms for decisions assisted by AI. Who answers when an AI-supported choice goes wrong is a question the law has not fully settled.

What this means for the market

Taken together, Japan’s regulatory picture is neither a barrier nor a free pass. It is a structured environment with clear classification, a specialist review team, a staged route to approval, and a mechanism for post-approval improvement. Those are real advantages for companies with strong clinical evidence and a precise intended use.

They are also a signal about where the weight of effort belongs. The paperwork is rarely the hardest part. Building convincing evidence, choosing the right claim, planning for reimbursement, and preparing for change after launch are the tasks that separate approved products from stalled ones. Companies that approach the PMDA early, treat consultation as a design tool rather than a formality, and collect Japanese data from the start tend to move with fewer surprises.

The road ahead will keep shifting. Generative AI, liability, and evidence transparency are all active questions, and the DASH for SaMD 2 goals run through the middle of this decade. Developers, hospital leaders, and investors who follow MHLW and PMDA announcements closely will be better placed to adapt than those who treat today’s rules as permanent.

Japan has already shown that it can regulate medical AI without freezing it in place. The next test is whether that same flexibility can extend to systems that write, reason, and respond. The answer will matter well beyond Japan’s borders.