Can AI Explain Your Blood Test Results? What It Can and Cannot Do

"Your ferritin is 14." Most lab reports tell you the number, print a reference range beside it, and stop there. The question people actually have — is this something to worry about? — is left unanswered, and that gap is why so many people paste their results into a search engine or a chatbot on the drive home.

So it is worth being precise about what software can and cannot do with a blood test.

What software genuinely does well

Putting a number in context. A result of 14 ng/mL means nothing on its own. Software is good at pairing a value with its reference range, your previous results, and the units your lab used — which vary between countries and are a common source of confusion when comparing results from different labs.

Tracking change over time. A single ferritin reading is a snapshot. Three readings across eighteen months is a trend, and the trend is usually the more informative thing. Keeping results in one place is unglamorous and genuinely useful.

Translating vocabulary. "MCV", "eGFR", "ALT" — a glossary is a legitimate, low-risk thing for software to provide.

Where it goes wrong

A reference range is not a threshold for health. Reference ranges are typically built from the middle 95% of a reference population, which means roughly one healthy person in twenty falls outside the range on any given test. Run a panel of twenty markers and it is more likely than not that at least one comes back flagged in a person with nothing wrong. Software that treats every out-of-range value as a finding will alarm people for statistical reasons alone.

Ranges are not universal. They differ by sex, age, pregnancy, ethnicity and by the analyser the lab used. A tool applying one set of ranges to everyone will misread results for a large share of the people using it — and the group most often mishandled is women, because male ranges are frequently treated as the default.

Language models can be confidently wrong. A system that generates fluent text about your results can produce a fluent, plausible, incorrect explanation, and it will not sound any less certain when it does. This is a well-documented failure mode, and it matters more in medicine than almost anywhere else.

A number is not a diagnosis. The same low haemoglobin is a different problem in a menstruating 25-year-old, a 70-year-old man, and someone recovering from surgery. Context that lives in your history and your symptoms — not in the PDF — is usually what decides the answer.

What this means for you

Use software to organise and to understand vocabulary and trends. Use it to arrive at your appointment with better questions. Do not use it to decide whether to worry, to change a medication, or to skip seeing someone.

And if a result is flagged, the useful next question is rarely "what does the internet say" — it is "compared to what?" Compared to your last result, to the range for your sex and age, to how you actually feel.

Where FeatherDoc is today

We should be straight about this, because the page you are reading sits on our own site: FeatherDoc does not explain your results yet. Today it does one thing — it stores the report you upload, encrypted and linked to your account, so it is in one place when you need it and you can delete it whenever you want. Nothing reads the contents: no text extraction, no AI, no interpretation.

We are building the explanation part next, and we would rather ship it late than ship the version described above — the one that flags every out-of-range value, applies one set of ranges to everybody, and sounds certain while being wrong. Everyone with a stored report gets an email when theirs is ready.