Skip to main content
International Standards

How To Add An AI Feature To A Medical Device That Is Already Certified

Adding AI to a certified device is a change to assess, not a new start. What the EU MDR, the AI Act and FDA change control plans expect, step by step.

Dario Esposito··6 min read
Hand touching a holographic AI interface with data panels, in teal tones

Adding an AI feature to a certified device is a design change that you must assess before release. In the EU, if it affects safety, performance or intended purpose, you notify your notified body and may need a new assessment, and the AI part brings AI Act obligations from August 2028 for devices that need a notified body. In the US, you assess whether the change needs a new submission, and you can use a Predetermined Change Control Plan (PCCP) to agree future model updates with the FDA in advance.

You do not start the certification from zero. You extend the evidence you already have.

Step 1: Decide whether the AI changes the intended purpose or the class

Start from the intended purpose statement. An AI feature that only speeds up an existing function is a different case from one that adds a new diagnostic or therapeutic output.

In the EU, software classification follows Rule 11 of the Medical Device Regulation (MDR): software that provides information used for diagnosis or therapy decisions is at least class IIa, and higher where a wrong decision could cause serious harm. A new AI output can move the device to a higher class.

Step 2: Assess the change under your current certification

  • EU: changes to design or intended purpose that affect safety or performance must be notified to your notified body, which may assess them. Your change-notification agreement with the notified body sets the thresholds. We recommend keeping a short, signed change-significance decision record that refers to the risk file.
  • US: assess the change against FDA's guidance on when a software change to an existing device needs a new 510(k). A new AI function often does.

Step 3: Plan for the AI Act (EU)

Most AI-enabled medical devices are "high-risk" under the EU AI Act, because they are covered by the MDR and need a notified body. After the AI Omnibus (Regulation (EU) 2026/1744), these obligations apply from 2 August 2028.

Joint guidance MDCG 2025-6 says you prepare one technical documentation file and one conformity assessment through the MDR route, with the AI Act requirements built in. Your notified body must also meet the AI Act requirements for notified bodies. Requirements to plan for: risk management across the AI life cycle, data governance for training and test data, automatic logging, transparency, human oversight, and declared accuracy and robustness.

A proposed MDR revision (December 2025) could change how the AI Act applies to devices, but it is not law. Plan for the rules in force.

Two more AI Act dates matter. If your AI feature generates text, audio, images or video, the Article 50(2) transparency rules already apply: systems placed on the market before 2 August 2026 must mark their output as AI-generated by 2 December 2026, when the grace period ends. And from 2 December 2027, high-risk obligations apply to stand-alone AI systems listed in Annex III, such as emergency triage systems that are not medical devices.

Step 4: Plan future model updates in advance (US PCCP, EU AI Act)

AI models change after release. A PCCP lets you describe planned modifications in your marketing submission and, once authorised, make them without a new submission each time, as long as you follow the plan. It has three parts:

  1. a description of the planned modifications;
  2. a modification protocol: how each change is made, tested and released;
  3. an impact assessment of benefits and risks.

On October 20th, 2026, the FDA published two hypothetical PCCP examples for machine-learning software: a sepsis alert adapted locally in each hospital, and an arrhythmia detection model retrained periodically.

In the EU, the AI Act has a similar rule. Changes to a learning AI system that you define in advance and document in the technical file are not a "substantial modification" (Article 43(4)), so they don't need a new conformity assessment. For MDR devices this applies from 2 August 2028. Until then, and for the MDR side, agree planned changes with your notified body.

Step 5: Extend the evidence, do not rebuild it

What you add to the existing file:

  • new and changed requirements, with their risks and controls;
  • data management records for training, validation and test sets;
  • verification and validation of the AI feature, linked to those requirements;
  • an updated risk analysis under ISO 14971 (the risk management standard);
  • updated post-market surveillance for model performance.

The work is easier when requirements, risks and tests are already linked, because you can see what the AI feature touches.

How P4SaMD supports this

P4SaMD, the compliance platform for Software as a Medical Device (SaMD) by Mia-Care, a Mia-Platform company, keeps requirements, risks, change requests and tests linked with two-way traceability, manages versions in line with IEC 62304 (the medical device software life-cycle standard), and uses a risk matrix aligned with ISO 14971. That gives you the impact view a change assessment needs, including AI risks. For the AI feature, P4SaMD is adding capabilities along the AI life cycle: registering and classifying each AI system, checking training data, blocking a model version that misses its acceptance criteria, and tracking models in production against the agreed change plan. P4SaMD supports compliance; the notified body or the FDA decides.

Want to see it in practice? On November 30th, Mia-Care hosts the free webinar "From Algorithm to Clinic: AI Medical Software to Market", with a case study by Aindo: mia-care.io/webinar/ai-medical-software.

FAQ

1. Do we need to recertify a device to add AI?

Not necessarily from scratch. You assess the change; if it affects safety, performance or intended purpose, the notified body (EU) or the FDA (US) may need to review it.

2. When do AI Act rules apply to AI medical devices?

From 2 August 2028 for devices covered by the MDR or IVDR that need a notified body. If the device generates text, audio, images or video, the Article 50(2) transparency rules already apply: systems placed on the market before 2 August 2026 must mark their output as AI-generated by 2 December 2026.

3. Do we need a separate AI Act technical file?

No. MDCG 2025-6 describes one technical documentation file via the MDR route.

4. What is a PCCP?

An FDA mechanism to pre-authorise a defined set of future changes, such as model retraining, in the original submission.