Template: AI-Enabled Device Transparency Information Sheet (Model Facts)

This is a ready-to-use transparency information sheet, sometimes called a model facts label, for an AI-enabled device or software as a medical device. A clinician cannot use an AI output safely without knowing what the model does, how well it does it, on whom it was validated, and where it fails. Transparency is also the control that makes meaningful human oversight possible; without it, oversight is a rubber stamp. Design this content into the device, the interface, the labeling, and the training rather than treating it as a single label paragraph. Replace every <<FILL: ...>> placeholder. A worked filled specimen follows. This is general guidance to adapt, not legal or regulatory advice.

Control header

The information below should be delivered in a form that fits the audience and the context of use; a specialist and a generalist may need different presentations of the same facts.

1. What this device does, and does not do

• Does: <<FILL: e.g. flags studies with suspected [finding] to raise their priority in the worklist (triage)>>.
• Does not: <<FILL: e.g. does not diagnose; does not remove any study from the worklist; the [clinician role] remains the decision-maker>>.
• Role in the workflow: <<FILL: an aid, not a replacement; where the output appears; what action it prompts>>.

2. How it performs

State validated performance with uncertainty, broken out by relevant subgroup. A single aggregate number is not transparency.

3. Population and conditions it was validated on

So a user can judge whether their patient is inside or outside the validated envelope.

• Patient population: <<FILL: ages, conditions, inclusion/exclusion>>.
• Sites and equipment: <<FILL: number of sites, scanner/analyzer vendors, settings>>.
• Conditions of use tested: <<FILL: workflow, image/assay types, realistic use conditions>>.

4. Known limitations and failure modes

State these plainly and usefully, not defensively for liability.

• <<FILL: e.g. lower sensitivity for small [findings] under [size]>>.
• <<FILL: e.g. under-represented in training: [scanner vendor / population], monitor accordingly>>.
• <<FILL: e.g. not validated for [population/condition]; do not rely on the output there>>.

5. Active version and change history

So that after any update no one in the field is unsure which model produced a given result.

Where the device has a predetermined change control plan (PCCP), each in-scope update refreshes the performance information and the active-version entry here.

6. How to use the output safely (human oversight)

• <<FILL: how to interpret a positive/negative or a score; what it does and does not license>>.
• <<FILL: when to disregard the output; the clinician retains final judgment>>.
• <<FILL: how to report a suspected error or unexpected behavior>>.

7. References

Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles (jointly published by FDA, Health Canada, and the MHRA, June 2024). Good Machine Learning Practice guiding principles (October 2021): provide clear, essential information to users. For a high-risk AI device sold in the EU, the transparency and instruction-for-use obligations under the applicable AI Act and MDR/IVDR requirements.

Confirm the current version of each reference before issue.

Acceptance criterion

A user can determine, from the materials provided, what the device does, how well it performs (with uncertainty and by subgroup), the population it was validated on, where it fails, and which version they are running, without contacting the manufacturer.

Filled specimen (excerpt)

Illustrative model facts for a locked chest X-ray pneumothorax triage SaMD.

What it does / does not: Flags chest X-rays with suspected pneumothorax to raise their worklist priority (triage). It does not diagnose, does not remove any study from the worklist, and the reading radiologist remains the decision-maker.

How it performs (excerpt):

Validated population: adults 18 and over, 9 hospital sites, 3 scanner vendors, upright and supine views. Not validated for pediatric patients or portable ICU films outside the tested settings.

Active version: v1.3, effective <<FILL: date>>; changed from v1.2 by recalibrating the triage threshold within the authorized range (PCCP modification M2); sensitivity unchanged, specificity improved by 1 point.

The specimen states the two limitations that matter (vendor B and small pneumothoraces) plainly, gives uncertainty on every number, and names the active version and what changed, so a clinician can judge trust and oversight without calling the manufacturer.

Common findings this sheet prevents

• A single aggregate accuracy figure with no subgroup breakdown and no uncertainty.
• Limitations written defensively for liability rather than usefully for the clinician.
• No version information, so after a PCCP update no one knows which model produced a result.
• Transparency treated as one label paragraph rather than designed into the device and interface.

How to adapt this template

1. Fit the presentation to the audience and context of use; a specialist and a generalist may need different views.
2. Give uncertainty on every performance number and break out the subgroups that matter.
3. State limitations usefully, not defensively.
4. Keep the active-version entry current at every in-scope PCCP update.
5. Confirm references against the current source before issue.