Software as a Medical Device (SaMD) regulation in Japan is administered by the Pharmaceuticals and Medical Devices Agency (PMDA) under the Act on Securing Quality, Efficacy and Safety of Products Including Pharmaceuticals and Medical Devices (the PMD Act, formerly the Pharmaceutical Affairs Law). For companies developing AI-based clinical decision support tools — including AI-assisted radiation therapy planning software — understanding how PMDA categorizes, evaluates, and approves SaMD is prerequisite knowledge before any clinical commercialization planning in Japan can be taken seriously.
This article describes the PMDA SaMD pathway framework as it applies to AI-based medical software. It does not constitute legal or regulatory advice; companies seeking PMDA registration for specific products should consult a Japan-registered regulatory affairs specialist and engage PMDA's consultation services directly.
How Japan's Medical Device Classification Works for Software
Japan's medical device classification system divides devices into four classes (I through IV) based on risk, broadly analogous to the FDA's Class I-III system but with different boundary conditions. Software medical devices — including SaMD — were explicitly incorporated into the classification framework through regulatory revisions effective in 2014, with further AI-specific guidance issued in subsequent years.
For AI-based clinical decision support software in radiation therapy, classification typically falls in Class II or Class III depending on the software's intended use and its impact on treatment decisions:
- Class II: Software that provides information to clinicians for review but where the clinician makes the final treatment decision independently, and where errors in the software output do not directly cause harm if the clinician applies normal professional judgment. Treatment planning decision support tools — where output is reviewed and approved by a qualified medical physicist before clinical use — often fall in this range.
- Class III: Software where output directly drives treatment delivery or where errors could cause serious patient harm without a separate clinical review step. Autonomous planning software without a mandatory physicist approval step would likely attract Class III scrutiny.
The boundary between Class II and Class III for AI planning software is not always clear-cut and is an active area of regulatory interpretation at PMDA. The software's intended use statement — precisely how the manufacturer characterizes what the software does and who makes the final clinical decision — is the key determinant. Manufacturers who frame their software as "physician/physicist-assisted decision support with mandatory human review" are articulating an intended use that points toward Class II; manufacturers who frame their software as "autonomous treatment planning" are describing Class III functionality.
PMDA vs. FDA SaMD Pathways: Key Differences
Developers who have previously navigated FDA's SaMD pathway under the 510(k) or De Novo process will find Japan's pathway structurally similar in some respects but distinct in important ones.
FDA's 510(k) pathway relies heavily on predicate device identification — demonstrating substantial equivalence to a legally marketed predecessor device. PMDA's Ninsho (approval) pathway for Class II devices is more documentation-intensive: it requires a detailed technical file including software lifecycle documentation, clinical performance data, and algorithm validation studies. PMDA does not have a direct equivalent to the 510(k) predicate-finding mechanism; each application is evaluated on its own clinical performance documentation.
For AI-based software specifically, PMDA has issued guidance on the evaluation points for machine learning-based medical devices, drawing on international frameworks including the IMDRF SaMD guidance documents and WHO guidance on AI in health. Key evaluation points include: the clinical dataset used for algorithm training and validation (diversity, representativeness for the intended patient population), the algorithm's performance on the validation dataset, and the change management protocol for software updates that may affect algorithm behavior.
This last point — change management for AI software — is where PMDA's approach adds significant ongoing compliance burden that is less prominent in pre-market regulatory frameworks. Because machine learning algorithms may be updated post-market, PMDA expects manufacturers to have documented change management procedures that determine when a software update requires a new conformity assessment versus when it can be implemented under a pre-approved change protocol.
The Consultation Process and Its Strategic Value
PMDA offers a formal consultation service (designated as "PMDA consultation") where companies can submit technical questions and regulatory strategy questions before filing a formal application. For AI medical device software, particularly in novel clinical applications like AI-assisted RT planning, pre-submission consultation is not merely available — it is strategically important.
PMDA's internal expertise in AI-based radiation therapy software is still developing, as is true of most regulatory agencies globally. Early consultation allows the manufacturer to educate PMDA reviewers on the technical specifics of the approach, receive early signals on documentation requirements that are not fully specified in published guidance, and avoid costly application revisions that result from PMDA identifying documentation gaps mid-review. Japanese regulatory consultants with PMDA experience consistently report that pre-submission consultation reduces total review time for novel AI applications compared to applications filed without prior consultation.
Clinical Performance Data Requirements
For AI-assisted RT planning software, PMDA will expect clinical performance data demonstrating that the software's output — in the context of its stated intended use — performs at a clinically acceptable level. What "clinically acceptable" means is defined by the manufacturer's intended use statement and supported by clinical validation data.
For a treatment planning decision support tool where the physicist reviews and approves all AI-generated plans, the clinical performance data framework typically includes: dosimetric comparison between AI-generated plans and plans generated by experienced physicists on the same patient cases; physicist reviewer assessment of plan acceptability (rate of plans accepted without modification, rate requiring modification, rate rejected); and, if available, outcome data from any prospective pilot studies at collaborating clinical sites.
We are not suggesting that PMDA requires randomized clinical trial data for SaMD approval — it does not, and requiring RCT-level evidence for treatment planning software would be disproportionate to the risk profile. The evidence standard is proportionate to device class and intended use. What PMDA does require is that the validation data is scientifically credible, the patient population is representative, and the performance metrics are defined and measured prospectively rather than selected post-hoc.
Post-Market Surveillance Obligations
Japanese medical device regulations require approved manufacturers to maintain a post-market surveillance system that collects adverse event reports, user complaints, and performance data from clinical use. For AI-based SaMD, post-market surveillance has an additional dimension: monitoring whether the algorithm's clinical performance in actual deployment is consistent with validation study performance, and whether any distribution shift in patient populations or clinical use patterns affects the system's outputs in clinically significant ways.
Practically, this means that the clinical infrastructure built during the validation phase — the methods for systematically comparing AI-generated plans against physicist assessments, the incident reporting pathways, the physicist training documentation — becomes the foundation of the post-market surveillance system. Manufacturers who build that infrastructure well during the approval process carry less regulatory burden in the post-market phase than those who treat validation as a documentation exercise.
For AI-based radiation therapy planning software in Japan, the regulatory pathway is navigable with appropriate preparation, realistic timelines — typically eighteen to thirty-six months from pre-submission consultation to approval depending on device class and preparation completeness — and a clinical validation program that is designed from the start with PMDA's evaluation framework in mind rather than reverse-engineered from existing data.