For readers and evaluation teams

Evaluate healthcare AI evidence

Follow a claim back to its source, compare what was measured and record the questions that remain. This worksheet organizes an evidence discussion. It does not score products or recommend clinical decisions.

Step 1

Read the claim in context

Start with How to evaluate a healthcare AI claim. Then use Compare healthcare AI evidence and its limits to distinguish a study design from a stronger or weaker answer to your particular question.

Step 2

Compare like with like

For a concrete example, read the AI scribe evaluation guide with the randomized trial evidence record. Compare the task, study setting and outcomes before drawing a conclusion.

An FDA AI device listing answers different questions from a study of outcomes in your setting.

Your evidence worksheet

Use public information only. Do not enter patient details, confidential records or passwords. Responses stay in this tab and are not submitted or saved by this tool. Closing or reloading the page clears them. Download a copy if you want to keep your work.

What specific task is claimed? Where and by whom would the tool be used?
Record a public study link, publication date and the exact claim being evaluated.
Who was studied? Who was excluded? How does the setting compare with the intended use?
What was the AI compared with: usual care, another tool, an expert or no comparison?
Record measured benefits and harms, denominators, follow up and uncertainty. Distinguish accuracy from patient outcomes.
Note design limits, missing results, conflicts of interest and questions about applicability.
Who should review the unresolved questions? What evidence would be needed before a decision?
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Check for missing fieldsDownload text copyPrint or save as PDFClear responses

Download the original blank evidence worksheet (PDF)

Fictional worked example

Invented for learning. This is not a real product, study or finding.

A fictional hospital evaluates a fictional scheduling assistant. Its claim is that staff spend less time drafting appointment reminders.

  • Intended use: draft reminders for staff review, not make clinical decisions.
  • Source: an invented internal report with no public study or independent verification.
  • Population: one scheduling team. Other sites, languages and patient groups were not evaluated.
  • Comparator: the team's usual reminder workflow.
  • Outcomes: the report describes drafting time but gives no denominator, error rate or uncertainty.
  • Limitations: selection methods, privacy arrangements and delivery failures are not reported.
  • Next discussion: ask the operational and privacy owners for the missing information before considering a local evaluation.

The missing outcome details prevent a reliable comparison. Completing the worksheet does not resolve that evidence gap.

Keep a useful record

Download your worksheet and record the public source date. Revisit the unresolved questions when new evidence becomes available. The tool does not retain your answers or send reminders.

For team implementation, continue to planning a responsible pilot. For visit preparation, use the patient learning path.

This original editorial tool is not a validated clinical instrument. The linked guides retain their own source and review information.