For teams building on MedGemma
Open medical weights make the draft cheap.
The signature is the part no model ships.
Google's MedGemma — 4B and 27B, text and image — is a strong starting point, and Google says so plainly: outputs are “not intended to directly inform clinical diagnosis, patient management decisions, treatment recommendations,” and “require independent verification, clinical correlation.” HarnessHealth is that verification, done by a named, NPI-verified physician, with a receipt anyone can check.
Not affiliated with Google. MedGemma is Google DeepMind's; the quotes are from its model page. Synthetic or de-identified outputs only until a BAA exists.
What the model's maker says about its outputs.
“MedGemma is not intended to be used without appropriate validation, adaptation and/or making meaningful modification by developers for their specific use case.”
Google DeepMind, MedGemma →“The outputs generated by these models are not intended to directly inform clinical diagnosis, patient management decisions, treatment recommendations, or any other direct clinical practice applications.”
Google DeepMind, MedGemma →“Performance benchmarks highlight baseline capabilities, but inaccurate model output is possible.”
Google DeepMind, MedGemma →“All model outputs should be considered preliminary and require independent verification, clinical correlation, and further investigation through established research and development methodologies.”
Google DeepMind, MedGemma →Seven use cases. Where the signature enters.
Google's seven potential MedGemma use cases, as it lists them, and the point in each where an output becomes a decision a named human has to own.
| Use case (Google's list) | What it does | Where a signature enters |
|---|---|---|
| High-dimensional medical imaging | Process data from CT, MRI, and histopathology scans | Upstream. A signature enters when the processed data becomes a report or a recommendation. |
| Longitudinal medical imaging | Assess chest X-rays against prior images to track temporal change | When the change assessment reaches a clinician or a chart, a named human owns it. |
| Anatomical localization | Localize anatomical features in chest X-rays | Upstream. No signature until it feeds a decision. |
| Medical document understanding | Extract structured data from medical lab reports | If the extraction feeds a claim, a determination, or a letter, the human who signs that document is the signature. |
| Medical image classification | Classify images across radiology, pathology, dermatology, ophthalmology | A classification that a patient or payer acts on is a clinical decision. Named human. |
| Medical image interpretation | Generate medical image reports and answer queries about them | A generated report is exactly the draft the gate holds until a physician signs it. |
| Medical question answering | Support for preclinical interviews, triaging, and clinical decisions | Triage and clinical decisions are where the state laws and WISeR now require a licensed human. |
Use-case names and descriptions from Google DeepMind's MedGemma page. The third column is ours.
How it wires in.
Your MedGemma app drafts
A report, an extraction, a triage note, an answer. Nothing about your model, your prompts, or your fine-tune changes. The output is tagged as a draft.
Draft-until-attested, in 90 seconds →A physician signs, revises, or declines
Specialty-matched, verified against the CMS NPPES registry, bound to a signing key only they hold. Outside a covered specialty the gate fails closed and tells you so.
Start in the sandbox →You get the receipt
NPI, timestamp, and document hash, anchored on hashcare.com and countersigned daily. Your compliance team verifies it without asking us — the independent verification Google asks for, made inspectable.
Verify a receipt →Straight answers
Does Google say MedGemma can be used clinically as-is?
No. Google states MedGemma is a starting point that is not intended to be used without validation, adaptation, or meaningful modification by developers, that its outputs are not intended to directly inform clinical diagnosis, patient-management decisions, or treatment recommendations, and that all outputs should be considered preliminary and require independent verification and clinical correlation.
Google DeepMind, MedGemma page →What does HarnessHealth add to a MedGemma application?
The accountable step Google says the outputs need: a named, NPI-verified, specialty-matched physician who signs, revises, or declines the output, and a hash-anchored receipt anyone can verify. We do not fine-tune, validate, or monitor MedGemma itself; we make the decision it feeds accountable.
How a pilot works →Is there evidence the human layer catches anything?
In the ClinicalSwipe lab note, MedGemma 27B and another open model ran about one hundred synthetic HSA/FSA determinations under the instruction never to invent a code. Every verdict was correct. Once, a model cited a correct ICD-10 code for a condition nobody had diagnosed; ninety re-runs at three temperatures never reproduced it. A failure that rare and that fluent is not caught by sampling — only by a named human on every determination.
The lab note, on the sandbox page →Bring the model you already chose.
MedGemma, Claude, GPT, Gemini, or your own weights. The gate does not care which model drafted it — only whether a licensed physician signed it.