Decoding a Medical Report With AI, Locally
How to locally break down terms and values from a medical report without sending the document to a cloud chat, and where you absolutely still need a doctor.
A medical report is often written for another doctor, not for the patient. Abbreviations, reference ranges, and cautiously worded phrasing have to be picked apart one at a time. AI can translate the text into plain language and help you prepare questions for your appointment. But uploading a document with a name, diagnosis, and test results into a random cloud chat is a bad price to pay for a convenient explanation.
A local model processes the text on your computer and never sends it to an AI provider.
What you can ask it to do
- Spell out abbreviations and explain medical terms in plain language.
- List values alongside the reference ranges given on the form.
- Build a timeline from several reports.
- Prepare a list of clarifying questions for your doctor.
- Flag which parts of the text the model couldn't read confidently.
This is help reading a document, not a diagnosis. The model doesn't see the patient, doesn't know the full medical history, and can be wrong even when it sounds confident.
Don't change treatment or dosages based on an AI's answer. For acute symptoms or urgent concerns, get medical help — don't wait for a model to go through the document.
How to break down a report locally
Install Doka, download a local model from the catalog, and create a separate working folder. Put a copy of the report in it — better to leave the original untouched. Then give it a task with explicit limits:
Explain this text in plain language. Don't diagnose anything or suggest treatment. For every conclusion, quote a short passage from the report. Flag anything unreadable separately.
If it's a scan or a photo, check the text recognition first. Pay especially close attention to dates, units of measure, decimal separators, and "above/below" markers. A single-character OCR error can completely change how a value gets interpreted.
Why "no cloud" isn't the same as full anonymity
With a local model, the text doesn't get sent to an outside AI service. But the file can still end up in a synced folder, a cloud backup, or a shared computer's history. Check where your working folder is stored, who has access to it, and whether disk encryption is turned on.
If the breakdown needs to go to a doctor, it's worth stripping unnecessary identifying details from the copy first. And for a clinic's actual workflow, one local laptop isn't enough — you need approved access rules, logging, and a proper medical information system.
How to check the model's answer
A good answer separates the document's text from the explanation. Ask for a table with four columns: "excerpt," "explanation," "what's unclear," "question for the doctor." Don't ask the model to guess a cause for an abnormal value with no context.
Take the reference range from the form itself: it depends on the lab, the method, age, and other conditions. Even a correctly transcribed value doesn't, on its own, say whether someone has a condition.
The bottom line
A local AI is useful as a translator from medical language into plain language, and as a helper before a consultation. It saves time, doesn't require sending the report to a cloud provider, and helps you put questions into words. The diagnosis and treatment decisions still belong to the doctor who knows the clinical picture.
If you want to go deeper into working with files, start with the article on AI document analysis.
Why locally, and not "any chatbot will do"
Medical documents are a case where the choice of tool isn't about convenience.
A report has everything at once: a diagnosis, a name, a date of birth, an insurance number, sometimes an address. Upload a file like that to a public service, and you're handing it over completely and permanently: even if the provider promises not to store requests, there's no way to verify that, and there's even less way to take it back once sent. And it's often not even your own data — it's a parent's or a child's.
A local model removes this question physically. The file opens on your computer, the request goes nowhere, because there's nowhere to send it — and "where did my test results end up" stops being a question at all.
Doka is built for exactly this scenario: the model downloads into the app and computes on the machine, PDFs and scans open where they live. Download it for free.
And the one thing no configuration changes: the model doesn't diagnose anything. It decodes terms, explains what abnormal values mean, and helps put questions into words for an appointment. The diagnosis and treatment decisions belong to the doctor who sees the full clinical picture.