IF MedTechStart a project

AI / ML

Models built on clinically labelled data, with the dataset, the training record and the performance evidence assembled the way a regulator will ask to see them.

When AI and Machine Learning Are the Right Step

When a product must interpret images or signals to screen, diagnose or monitor. The model is medical device software, so the data and the testing matter as much as the accuracy.

  • Read Medical Images

    Classify or measure findings in images such as scans, slides or photographs, with clinicians setting the reference.

  • Interpret Signals

    Detect patterns in signals such as ECG, breathing or motion, on the device or in the cloud.

  • Prepare a Model for Approval

    A model that works in the lab and now needs a documented dataset, a locked version and a test on unseen data.

How it works

How AI and ML Development Works

  1. Step 1

    Define the Task

    The clinical question, the population and the reference the model is judged by.

  2. Step 2

    Build the Dataset

    Collected and labelled with clinicians, then split into training, tuning and test sets.

  3. Step 3

    Train and Lock

    Models trained and compared, then one version locked for testing.

  4. Step 4

    Prove Performance

    Tested on data it has never seen, with a plan for how it may change.

What you get

What You Bring and What You Get

You Bring

  • The clinical question, and who the output is for
  • Data you already hold, and its consent terms
  • Access to clinicians for labelling, if you have it
  • The device or software the model runs in
  • Any earlier models and their results

You Get Back

  • Dataset description and labelling protocol
  • Model and training record
  • A locked model version for testing
  • Performance report on held-out data
  • Plan for future model changes
  • Inputs on model errors for the risk file
Questions

Questions Clients Ask

How much data do we need?

It depends on the task, how varied the patients are and the performance you must show. The test set must be large enough to estimate performance with useful confidence, and it should reflect the people the device will be used on.

Can the model keep learning after launch?

Not freely. A regulated model is usually locked, and a change is planned, tested and approved like any other design change. Some regulators accept a plan agreed in advance that sets which changes may be made and how they are tested.

Who labels the data?

Clinicians, working to a written labelling protocol. Where experts disagree, the protocol sets how the reference label is decided.

Have a question about AI / ML?

Opens the contact form with this service named. Nothing is sent until you press send.