Two regulatorily distinct kinds of AI are entering clinical settings: voice entry and document drafting in an EMR, and AI detecting lesions in images. Both get called "medical AI," but the former is generally not a medical device while the latter may constitute Software as a Medical Device (SaMD) under pharmaceutical and medical device law.
Disclaimer: This article provides general information. Medical device status is determined product by product, and regulation is subject to revision. Always verify with current primary sources (the Ministry of Health, Labour and Welfare, PMDA) and the responsible authorities.
What SaMD Is
SaMD is termed a program medical device under Japanese law: software that functions as a medical device in itself, rather than software embedded in hardware.
Under pharmaceutical and medical device law, programs meeting certain requirements fall within the definition of a medical device, bringing requirements for marketing authorization holder licensing and per-product approval, certification, or notification.
Status Is Not Determined by Developer Intent
The most important point. Whether software constitutes a medical device does not depend on whether its developer considers it one. It is determined by statutory definition.
Two perspectives govern:
- The degree of involvement in diagnosis, treatment, or prevention of disease
- The level of risk to life and health should a malfunction occur
The higher both are, the more likely the software qualifies—and the heavier the regulatory class.
Examples on Either Side
Generally not a medical device
- Programs that merely record, display, and store data such as weight and step count without processing it
- Programs providing general health information not based on an individual patient's condition
- Programs supporting work processes such as documentation, booking, and claim generation
Potentially a medical device
- Programs detecting candidate lesions in images to support diagnosis
- Programs processing test data to indicate disease likelihood
- Programs directly involved in determining treatment policy
The dividing line is broadly whether the software merely handles data or participates in medical judgment.
Classification of diagnostic support programs is evaluated on the significance of the diagnostic result, the risk of malfunction, and whether it supports or replaces the physician's final judgment.
Which Are EMR AI Features?
| Feature | Position |
|---|---|
| Generating chart entries from speech | Documentation support; generally workflow support |
| Searching and summarizing past records | Information organization; generally workflow support |
| Drafting referrals and certificates | Document support; generally workflow support |
| Checking for missed billing | Claims support; generally workflow support |
| Automating booking and reception | Workflow support |
| Detecting lesions in images | May constitute a medical device |
| Indicating disease likelihood from test values | May constitute a medical device |
Most functions an AI-native EMR provides fall under support for documentation, paperwork, claims, and booking. Where diagnostic support is placed within the same interface, that portion may fall under a different regulatory framework.
The functional scope is covered in What Can an AI-Native EMR Actually Do?.
Whether a specific product qualifies is a per-product determination. The above is a general orientation; actual determinations require confirmation with the responsible authority.
What Qualification Entails
- Marketing authorization holder licensing
- Per-product procedures: notification, certification, or approval depending on class
- Quality and safety management systems, including post-market malfunction reporting
- Advertising restrictions applicable to medical devices
Regulatory weight scales with classification—higher risk, heavier procedure.
Programs Promoting Development
DASH for SaMD. A joint package strategy from the Ministry of Health, Labour and Welfare and the Ministry of Economy, Trade and Industry. Under it, consultations on medical device status are handled centrally by the responsible ministry division. A second-phase strategy has also been published.
Two-stage approval. In November 2023, guidance established a staged approach reflecting SaMD characteristics: granting approval early once a degree of efficacy is confirmed, then advancing evaluation based on real-world use. Industry associations have published illustrative case collections.
PMDA capacity. In July 2024, PMDA's SaMD review function was reorganized from a division into a department with expanded capacity. PMDA also publishes a guide to regulatory development and approval applications for SaMD.
What Providers Should Verify
Is this function a medical device or workflow support? Ask the vendor explicitly. If a feature claims diagnostic support, confirm whether it holds approval or certification.
If approved, within what scope? Use must stay within the approved intended use, target, and conditions. Distinctions arise—usable for this examination, not contemplated for that one.
Reimbursement treatment. Holding medical device approval and being reimbursed are separate matters. If you are considering payback on cost, verify them separately.
Where responsibility for final judgment sits. Most important of all: even when using AI-based diagnostic or therapeutic support, the responsibility for final judgment rests with the physician. Adopting an AI output verbatim does not transfer responsibility to the AI.
See Where to Draw the Line on Delegating Work to AI.
The Relationship to Generative AI
Recent generative AI sits awkwardly within existing frameworks because output is probabilistic and the same input need not yield the same result.
Most generative AI currently deployed in clinical settings occupies workflow support—documentation, summarization, drafting—where the premise is that a human verifies and finalizes the output, keeping regulatory position and operational safety aligned.
Legal orientation is covered in Using Generative AI in Healthcare.
Conclusion
- SaMD is software functioning as a medical device in itself
- Status is set by statutory definition, not developer intent, judged on involvement in diagnosis and treatment and on malfunction risk
- Programs that only record, display, and store, and workflow support programs, are generally out of scope
- Lesion detection in images and disease likelihood from test values may qualify
- Most EMR AI features are workflow support, but adding diagnostic support changes the treatment of that portion
- Japan has DASH for SaMD, two-stage approval, and expanded PMDA review capacity as promotion measures
- Providers should separately verify device status, approved scope, reimbursement, and responsibility
- Even with AI support, responsibility for final judgment rests with the physician
For details on AI Karte or to request a demo, please contact us.