IF MedTechStart a project
A smartphone showing a glucose reading and its trend, beside a handheld meter on a desk.

AI and SaMD

Software that supports a clinical decision and is regulated as a device in its own right. The dataset, the validation protocol and the change-control plan matter as much as the model.

Devices we have built here
Dermatology screening platforms, tele-consultation platforms, image classification, clinical AI tools.
The challenge

What Makes These Devices Hard

The Dataset

Labelled by clinicians, representative of the people it will be used on.

Validation

A locked model tested on data it has never seen, to a written protocol.

Change Control

A plan for how the model is retrained and when that is a new submission.

Cybersecurity

Patient data protected in transit, at rest and in the cloud.

Before you start

Tell Us These First

Useful in a First Enquiry

  • The clinical question, and who acts on the output
  • The data you hold: size, source sites, devices and consent
  • How the labels were made, and by whom
  • Where it runs: phone, hospital system or cloud
  • Whether the model should keep learning after release
Your path

The Path for a Device Like Yours

  1. Step 1

    Idea Validation

    Define the clinical question, the user and the intended use the model will be claimed for.

    Idea validation
  2. Step 2

    AI / ML

    Build the dataset plan, labelling protocol and model, with training, tuning and test data kept apart.

    AI / ML
  3. Step 3

    Mobile / Web / Cloud App

    Build the application and cloud around the model under IEC 62304, with security designed in.

    Mobile / web / cloud app
  4. Step 4

    Verification & Validation

    Test the locked model on independent data to a written protocol, and test usability.

    Verification & validation
  5. Step 5

    International Regulatory

    Prepare the file for each market, including how model updates will be handled.

    International regulatory
  6. Step 6

    Post-Market Support

    Monitor real-world performance and feed drift and complaints into change control.

    Post-market support
Questions

Questions About AI and SaMD Devices

Can our model keep learning after release?

It can be planned, but each change to a regulated model needs control. Some regulators accept a plan agreed in advance that sets which changes are allowed and how each is verified; other changes need a new submission. Many first releases use a locked model, which keeps the evidence simpler.

Is our software a medical device?

If it is intended to diagnose, monitor or guide treatment, it is likely to be. The intended use decides, not the technology. Its class follows the intended use and the classification rules; try the class and route tool.

Is a retrospective dataset enough evidence?

A retrospective study on independent data is often the start. Whether prospective evidence is also needed depends on the risk, the claim and the regulator. Plan the evidence against the intended use before labelling begins.

Building a device in AI and SaMD?

Opens the contact form with the topic filled in. Nothing is sent until you press send.