Skip to the content
AI Test AutomationPreview AI Security Learning Center

AI Test AutomationAI health testingAI Health Test Lab

AI health testing Preview

AI Health Test Lab: a second opinion you can test

Lab results and medical images, explained in plain language by an AI reader, and the harness that checks the explanation: units, sides, privacy, and what it must never say. Run in your browser, nothing uploaded, every sample made up.

AI reads your lab results

Paste your numbers, get them back in plain language

Type them, paste them or import the file your lab gave you. Test names are read in English, Spanish, French, Portuguese, German, Italian and Japanese, with a decimal point or a decimal comma; a name in any other language can be matched by hand under the box. When the report prints the earlier result beside the new one, add it as prev 12.6, or import a table with a Previous column: the reader says which way the number moved. Nothing is kept between visits; it reads only what is on the report. One result per line, like Hemoglobin 13.4 g/dL (12.0-16.0). If your report prints its own range, put it in brackets: it always wins, because ranges differ from lab to lab. Other units are converted, and a test it does not know is said out loud, not dropped.

or drop a file on the box below

Reads text, CSV and other tables, JSON, FHIR, HL7, HTML, XML and PDFs that hold real text. The file is read here in your browser and goes nowhere.

The plain-language report

Not medical advice. This does not replace a medical professional: it is a test page that shows what AI can do with a lab report, and how to check it.

Ranges: Merck Manual: blood tests, reference ranges · MedlinePlus: cholesterol levels · MedlinePlus: triglycerides test · MedlinePlus: hemoglobin A1C test. They are general adult ranges and they differ between labs. The line for "well outside" is this lab's own rule: a value beyond the range by more than three quarters of its width, or by more than the limit itself for a one-sided range. It is a prompt to call, not a clinical threshold.

AI looks at a scan

A synthetic MRI, a patch the reader has to find, and everything it says about it

This is a made-up head scan, sixteen slices, with a bright patch planted on one side. The reader compares every slice with its own mirror image, marks what stands out and says so in words. Drag on the picture to change window and level, scrub the slices, or open your own DICOM files (they stay on your computer).

You can also drop .dcm files on the picture.

What the reader says

The file's tags

TagNameValue

Window and level is the linear mapping from DICOM PS3.3 C.11.2: VOI LUT module (window center and width). This reader decodes uncompressed images only; a compressed file shows its tags and says why it cannot show the picture. The patch finder is tuned to the synthetic head on this page, so it is switched off for your own scans: on a real study it would mark normal anatomy.

AI privacy check

A scan carries the patient's name in more places than the name field

A DICOM file stores the patient's name, ID and birth date beside the picture. The standard's confidentiality profile lists what to empty, remove or replace. This check does that, and then does what the list cannot: it looks for the name and ID written into free text, like a study description. The synthetic scan hides the patient's name in exactly such a place.

Found in the open file

    After de-identifying

    The picture is untouched; only the tags change. The copy says it was de-identified (Patient Identity Removed = YES).

    Actions follow the examples in Table E.1-1 of DICOM PS3.15 Annex E: Attribute Confidentiality Profiles: Patient's Name, Patient ID, Patient's Birth Date, Study Date and Referring Physician's Name are emptied; Patient's Address and Other Patient IDs are removed; the SOP Instance UID is replaced. The standard itself says following the profile does not by itself guarantee that all identifying information is gone, and text burned into the pixels is not touched here.

    AI grades the AI

    Paste any AI's explanation of the results above and see how it does

    Ask a chatbot to explain the results in the box at the top, paste what it says here, and the grader checks it against the numbers you gave: does every figure come from your results, does it name each result that is out of range, does it get high and low the right way round, does it say when something is urgent, does it stay out of diagnosis and treatment, and could a 9th grader read it.

      AI tests the reader

      38 tests, 18 ways to break the reader, and proof the tests notice

      Tick a defect and the reader above starts failing in that way: it forgets the sex, skips a unit conversion, tells you that you have a condition, leaves a patient's ID in the file. Run the suite and watch the right test go red. The mutation run does this for every defect in turn and counts how many the tests caught.

      The same files run in Node: node data/testlab/health/run.js runs the suite and node data/testlab/health/run.js --mutate runs the mutation check. Put them in a pipeline and a change to the reader that breaks a unit, a side or a privacy rule cannot ship quietly.

      AI testing, not AI showing off

      The reader on this page is a stand-in. The tests are the point.

      The reader here is rule-based, so it runs offline and gives the same answer every time. A real product would put a language model or an imaging model in its place. What does not change is what you must prove about it, and this page is how you would prove it.

      Ground truth you control

      The scan is synthetic, so the planted patch has a known place, side and size. The test asks whether the reader found that patch, on those slices, about that size, and nothing on a healthy scan, across five noise patterns.

      Rules that must always hold

      Units are converted before values are compared. A printed range beats a textbook range. A value on a limit is inside it. High is never called low. These are cheap to test and costly to get wrong.

      Limits on what it may say

      No numbers that were not in the results. No "you have". No medication instructions. No "no need to see a doctor". Always a pointer to a clinician. Plain enough for a 9th grader. These are checks on language, and they run on any model's output.

      Privacy as a test

      The test opens the de-identified file and looks for the name, the ID and the birth date anywhere in it, including free text, and checks the picture did not change. A leak is a failing test, not a surprise in an audit.

      Tests that are tested

      Eighteen planted defects, each caught by at least one test. If a defect gets through, the suite has a hole, and the mutation run shows exactly where.

      Plug in a real model

      Ask your model to explain the results above, paste its answer into the grader, or call it from the same suite in your own pipeline. This page never sends health data anywhere, so it does not call a model for you.

      Sources

      Read on these pages, 2026-10-31

      Unit conversions are the standard factors (glucose mmol/L times 18.016, creatinine micromol/L divided by 88.4, hemoglobin g/L divided by 10). The code is data/testlab/health/labs.js, dicom.js and health_tests.js; the page runs them as they are. MedlinePlus notes that reference ranges vary with the lab, age and sex, which is why the range printed on a report is used first. This is education, not medical advice.