Skip to content

Entering Blood Test Results Online for Free: What the Tools Do and What to Build Instead

Oak
A laboratory machine feeding cloudy film sheets into a glowing grid of small glass ampoules that assembles behind glass.

You can enter blood test results online free in dozens of places, and nearly all of them do the same thing. They extract the analyte names and numbers from your PDF, compare each value to a generic reference interval, and ask a language model to write a paragraph about anything flagged high or low. That is useful once. It is close to worthless the third time, because the output is not stored in a form you can compute on, and a single value compared to a population range tells you much less than four values of the same analyte compared to each other. The higher-value move is to spend an hour building a small structured store of your own results, keyed by LOINC code and timestamp, and then point whatever model you like at that. Nothing in this piece is diagnosis, and no free tool should offer you one either.

What the free web analyzers give you

The underlying mechanic in the popular tools is optical character recognition or PDF text extraction, followed by a prompt. The interpretation quality is roughly the quality of the model behind it, which means it is decent at explaining what alanine aminotransferase measures and poor at anything requiring your history. These tools are also stateless by design. You upload, you read, and you close the tab. The parsed values are then gone.

They share three failure modes worth knowing about. First, flag columns: labs print an “H” or “L” or an asterisk next to out-of-range values, and extractors frequently attach that character to the numeric field or drop the reference interval entirely. Second, reflex and add-on tests, where a panel triggers a follow-up assay reported in a separate block with its own units. Third, unit ambiguity, which is the one that causes real misreadings. We cover it below.

There is also a structural problem. A generic reference interval is the central 95% of a reference population, usually stratified only by sex and sometimes age. It is not a statement about you. A person can sit inside the interval for every analyte while moving steadily across it over five years, and no single-timepoint tool will notice.

The schema we would use

We would keep one long-format table, one row per measurement, and nothing else. Long format survives new analytes without schema migration. A CSV or a single SQLite file is enough. Columns:

collected_at     ISO 8601 with timezone, e.g. 2026-03-14T07:42:00-07:00
analyte          human label, e.g. "Alanine aminotransferase"
loinc            e.g. 1742-6
value_raw        numeric as reported
unit_raw         as reported, e.g. "U/L"
value_si         converted
unit_si          canonical
ref_low          as printed on that report
ref_high         as printed on that report
lab              Quest, Labcorp, hospital, etc.
method           assay/platform if printed
fasting          boolean or null
specimen         serum, plasma, whole blood
notes            hemolysis comment, "calculated", etc.

Keying on LOINC matters because it is what makes results from different labs comparable without string matching. LOINC codes distinguish the same analyte measured on different specimens and in different units, which is exactly the distinction that breaks naive joins. Look codes up in the LOINC search rather than trusting any code from memory, including ours. Storing both the raw and converted value is cheap insurance against a bad conversion factor.

Record the printed reference interval per report. Do not record it once globally. Intervals differ between labs and change when a lab switches platforms. When you later see a step change in a series, the first thing to check is whether the lab or method changed under you.

Getting numbers out of the PDF without transcription errors

Manual entry of a full metabolic panel plus a complete blood count is about forty numbers, and hand-typed rows carry real error rates. We would rather extract and then verify.

For text-layer PDFs from Quest or Labcorp, pdftotext -layout report.pdf - preserves column structure well enough that a short regex pass gets most rows. For scanned reports, ocrmypdf --force-ocr first, then the same extraction. A multimodal model reading the page image is now the fastest route for messy layouts, and we use it with one rule. Require the model to emit strict JSON containing the value, the unit, and the printed reference interval. Then reject any row where the value falls outside its own printed interval without a corresponding flag in the source. That single consistency check catches most decimal-point and digit-transposition errors.

If your results live in a patient portal, check for a FHIR export before scraping anything. Many US portals expose an Observation endpoint under a patient-facing API. Each Observation already carries a LOINC code in code.coding, a valueQuantity with UCUM units, and a referenceRange. That maps onto the schema above almost directly. Apple Health can also export clinical records as XML, though coverage depends on your provider.

Units, and the conversions that matter

Unit confusion is the most common way a self-interpreted lab panel goes wrong, especially when mixing US and non-US reports. Convert everything once, store it, and never reason over mixed units.

glucose        mg/dL x 0.0555 = mmol/L
total chol     mg/dL x 0.0259 = mmol/L
triglycerides  mg/dL x 0.0113 = mmol/L
creatinine     mg/dL x 88.4   = umol/L
hemoglobin     g/dL  x 10     = g/L

Two traps. LDL cholesterol on a standard lipid panel is usually calculated, not measured, and the equation used matters at high triglycerides. Estimated glomerular filtration rate is a formula output, not an assay, so a change in eGFR can reflect an equation change rather than any change in you. Keep calculated quantities marked as such in the notes field.

Deciding whether a change between two draws is real

This is the analysis the free tools never do, and it is the one that makes a longitudinal table worth building. Every result carries analytical imprecision from the assay (CVA) and within-person biological variation across time (CVI). The reference change value tells you how large a difference must be before it is unlikely to be noise:

RCV = 1.41 x Z x sqrt(CVA^2 + CVI^2)

Z is 1.96 for a two-sided 95% criterion. Published within-person variation estimates live in the EFLM Biological Variation Database, and they vary enormously by analyte. Hemoglobin A1c is tight, with CVI on the order of a couple of percent, so a small move is meaningful. Alanine aminotransferase and thyroid-stimulating hormone have CVI in the range of roughly 20%, so two results differing by a third can be entirely consistent with no change at all. Compute the RCV per analyte, store it as a column in a lookup table, and have your tooling label each delta as within noise or beyond it.

Once you have four or more timepoints, prefer a plotted series with a fitted slope over any pairwise comparison. A slow drift that never leaves the reference interval is the pattern worth seeing, and it only appears in the series.

Handing the table to an AI agent

A model reasoning over a clean long-format table plus RCV annotations behaves very differently from one reading a PDF. Give it the whole series, the printed intervals, and the lab and method columns. Include the fasting flag. Ask it to identify analytes whose deltas exceed RCV and analytes with monotonic trends, and ask it to state which apparent changes coincide with a lab or method switch. Instruct it to return the arithmetic so you can check it.

Keep the model away from two things. Do not ask it for a diagnosis, because a panel of twenty analytes with a 95% interval each will produce out-of-range values by chance alone, and the model will happily narrate them. Do not ask it about medication changes. Any result that is flagged and unexpected, any value far outside its interval, and any trend you find concerning belongs in front of a physician who can take a history and order confirmatory testing. Repeat testing on a fresh draw is the standard first step for a surprising single value, and ordering that is a clinician’s call.

Questions people also ask

Can I enter blood test results online and get a diagnosis?

No. These tools produce explanatory text, not a diagnosis, and treating their output as one is the main risk in using them. Diagnosis requires history, examination, and often confirmatory testing. That means a licensed clinician.

Is a free AI lab interpreter accurate?

The explanation of what an analyte measures is usually accurate. The extraction step is where errors enter. They come in through flag characters, reflex tests reported separately, and OCR digit errors. Verify every number against the source PDF before you act on any reading.

What is the best free way to store my own results over time?

A single CSV or SQLite table in long format, keyed by LOINC code and collection timestamp, with the printed reference interval and lab recorded per row. Plot each analyte as a series. That costs nothing and is more informative than any one-off upload.

Why do my reference ranges differ between labs?

Reference intervals are derived from each lab’s own reference population and assay platform, so they shift when a lab changes methods. This is why you store the interval that was printed on each individual report rather than a single global range.

Two points give you a delta you can test against the reference change value for that analyte. Four or more give you a slope worth looking at. Below that, most apparent movement is analytical and biological noise.

Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.