AI Regulation Tracker  /  Regulator framework

Japan routes self-updating medical AI through IDATEN pre-agreed change plans

Under the Pharmaceuticals and Medical Devices Act, continuously-learning diagnostic software can ship in Japan only with a change-management plan locked in at approval and ongoing performance monitoring. The supervising physician still owns the final call.

By Anthony Guerriero  ·  Reviewed by The Leveraged Years Editorial Desk  ·  Published July 7, 2026  ·  Last updated July 7, 2026

Japan will let a medical AI system rewrite itself in the field, but only if the maker agreed in advance to exactly how. That is the core of IDATEN, the 変更計画確認手続制度, a pre-approved change-plan confirmation system that governs continuously-learning software-as-a-medical-device under the Pharmaceuticals and Medical Devices Act. Rather than treat every model update as a fresh device requiring a fresh review, IDATEN lets a manufacturer file a change-management plan at initial approval. Changes that fall inside that plan then move through a fast-tracked partial-change approval instead of a full new application.

The regulators are PMDA, the Pharmaceuticals and Medical Devices Agency, working with the Ministry of Health, Labour and Welfare. Their bargain is straightforward. A developer gets speed and predictability for the updates it can describe up front. In return, it accepts a binding scope: the plan defines the permitted envelope, and modifications outside it do not inherit the fast-track route.

What IDATEN actually requires

Two obligations sit at the center. First, the change-management plan must be part of the approval package, not a later add-on. It commits the manufacturer to a defined set of anticipated changes and the methods used to validate them. Second, for continuous learners, performance monitoring is mandated. A model that keeps learning from new data can drift, and IDATEN treats ongoing post-market monitoring as the price of allowing that learning to happen without a full re-review each time.

Whether a given piece of software counts as a regulated device in the first place turns on its role. Under MHLW's program-applicability approach, medical-device status is judged on how much the software contributes to treatment or diagnosis and the risk posed if it malfunctions. Low-stakes reference tools sit outside the regime; software that materially drives a diagnostic or treatment decision sits inside it.

The practical effect is a shift in where a company's regulatory effort goes. Under a conventional regime, each retrained version of a learning model would arrive at the regulator as a separate event, each with its own review clock. IDATEN front-loads that work. The manufacturer spends more at the design and submission stage describing the anticipated changes and the validation methods that will police them, and less on repeated full filings afterward. The reward is throughput: a firm that can specify its update path can iterate inside the confirmed envelope at partial-change speed rather than waiting on a new approval for each improvement.

The physician still decides

IDATEN does not move clinical responsibility onto the algorithm. Diagnostic-support tools inform the physician; they do not replace the physician. The supervising doctor retains the final-decision and oversight duty. For hospitals, that means an approved self-updating tool is a second opinion with a regulatory pedigree, not a delegation of liability. The clinician who acts on the output still owns the outcome.

What it does not do

IDATEN is not a blanket license for autonomous, unmonitored learning. It does not waive performance monitoring, it does not cover changes a manufacturer failed to describe in its plan, and it does not shift the diagnostic decision away from the treating physician. It is a controlled lane for pre-agreed evolution, not an exemption from oversight. A change the maker did not anticipate still needs the normal approval path.

The signal from May 2025

In May 2025 PMDA began publishing usage-performance information, a move toward making the real-world behavior of these products visible rather than leaving it inside company files. For a market that is allowing software to update itself in clinical use, published performance data is the accountability layer that keeps the arrangement honest.

For a US reader, the cross-border point is direct. IDATEN is the close analogue of the FDA Predetermined Change Control Plan, or PCCP, the mechanism the FDA uses to let AI and machine-learning devices update within a pre-cleared plan. A manufacturer selling in both markets is answering the same question twice: what will this model be allowed to change on its own, and how will you prove it still works. Building the IDATEN-to-PCCP mapping early lets a company design one change-management story that satisfies both regulators instead of two disconnected submissions. The alternative, treating Japan and the US as unrelated filings, tends to produce two versions of the same monitoring plan that then have to be reconciled anyway.

Browse the full AI Regulation News tracker

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.