Blood biomarker analysis, from value to sourced reading.

In four steps, each of which can be inspected. Parse: the result arrives as a lab PDF, a photo, an HL7 or FHIR feed or JSON, and the value, unit, reference range and flag are computed by deterministic code. Structure: the biomarker is mapped to one of 17,000+ LOINC codes and its unit to UCUM, and it becomes an Observation in one FHIR R4 record for that person. Understand: the biomarker is analysed against its own history, the rest of the panel and the record, by several models that cross-check one another, with named guidance cited claim by claim. Deliver: a clinician view, a patient view, a PDF, or JSON through the API.

Because a number in a medical analysis has to be right on every run, and a language model is not the tool for that. Converting a unit, comparing a value with a range and deciding whether it is flagged are things code does the same way every time. Vitrubo does them in code. A scanned PDF or a photo is recognised by several models in parallel and their readings are cross-checked before a value is accepted, and the reasoning the models then write is checked back against the finished numbers. No number in the analysis is written by a model, and the flag follows the laboratory's reference range, not a threshold a model chose.

Only after normalisation. Laboratories name the same blood biomarker differently, report it in different units and publish their own reference ranges. Vitrubo maps each biomarker to one of 17,000+ LOINC codes and converts each unit to UCUM, so a value from one lab in one year and a value from another lab the next are analysed as points on one line. The reference range is not normalised away: it is kept per laboratory, because the range that applies to a value is the one the lab that measured it published. That is what lets a trend be read across laboratories without pretending the laboratories were the same.

A biomarker on one date is either inside or outside its range. The same biomarker on several dates has a direction, and the direction often says more than the latest value. A value that has crept up across a year while staying inside its range, a value that fell back into range after a period outside it, a value that has never moved: each is read differently, and the analysis says which it is. It also says how old the latest measurement is. A biomarker last measured long ago is marked as stale, and where the guidance asks for a repeat, the re-test is named beside the value. The trend is read on the normalised value, so a change of laboratory between two draws does not show up as a change in the biomarker.

Because biomarkers move together, and the meaning of one often sits in another. A borderline result can matter because of the result beside it; a flagged result can be expected because of a medication on file; a set of values each just inside its range can point in one direction when read together. Vitrubo analyses each blood biomarker alongside the others in the panel and alongside the conditions, medications and history in the record. The output is a few connected stories rather than a list of isolated flags, and each story names the biomarkers it rests on.

Several models analyse the same case independently and their outputs are compared. Where one model states a finding another does not, the omission is caught and filled; where two models read a biomarker differently, the disagreement is resolved with reference to the numbers and the guidance rather than averaged into a hedge. The draft the clinician sees is the result of that cross-check, not the first answer of one model. On Stanford MedAgentBench, Vitrubo completed 600/600 tasks across five runs.

From named publications: EHA, WHO, NICE, KDIGO and BSH among others, with up to 30 sources behind one biomarker view. Every statement in the analysis carries a chip, and the chip opens the publication the statement rests on. That is the difference between an analysis and an opinion: the clinician who signs can read the basis, agree with it or overrule it. Nothing in a Vitrubo blood biomarker analysis is asserted without a source that can be opened.

Into connected stories placed in four zones: needs attention, worth watching, back to normal, always normal. A biomarker that returned to its range since the last draw is named as such rather than quietly leaving the list, and a biomarker that has never moved is said to be steady rather than omitted. Each story carries the biomarkers behind it, the trend of each across draws, and the guidance behind the interpretation. Grouping is part of the analysis rather than a layout choice: a story exists because the biomarkers in it were read together, against the same guidance. The zones are the same in the clinician view and the patient view; what changes is the language.

They are two views of one analysis, not two analyses. The clinician view carries the full reasoning: how each biomarker was read, the conditions and medications it connects to, the sources, and the tests worth considering next. The patient view carries the same reading in plain language: which biomarkers stand out, how the stories fit together, and what to raise with a doctor. Both regenerate together whenever a new result arrives. Vitrudoc, an AI chat grounded only in the person's own record, can sit beside either view to answer questions about it.

Yes. Every new result regenerates the analysis of the whole record, not only of the new values: an earlier biomarker can move from worth watching to back to normal because of what the new draw shows, and a value that was steady can become a trend. Earlier versions of the analysis are kept as history, so the reading made at a previous visit can be revisited after the next one. A clinician comparing two versions sees which biomarker moved the reading and why. The report is a living reading of the record rather than a page written once and filed.

No. Vitrubo is not a diagnostic device, does not select or recommend treatment, and does not replace the clinician or laboratory scientist. It drafts an analysis of the biomarkers that are there, organises the record, and names the tests worth considering; the professional reads the draft and signs it, or does not. For a person reading their own results, the analysis is educational context to bring to an appointment, not medical advice. Analysis here means the reading of measured values in their context, not a clinical judgement about the person.

The laboratory keeps measuring as it does now. Results reach Vitrubo from the LIS as HL7 or FHIR R4; each flagged biomarker is analysed against the laboratory's own reference range, the patient's earlier results where they are held, and the guidance for that marker; and a drafted comment sits beside the result. The lab scientist approves, edits or discards the comment at release. Where the analysis suggests a follow-up test, it is raised while the sample is still fresh rather than after the report has gone out. The comment is drafted in the laboratory's own style, configured during the evaluation.

On its own panels, starting with a call, then sandbox access and API documentation. The team sends anonymised biomarker results in its own formats and receives the FHIR R4 record and the analysis as JSON. Its medical team reads each drafted analysis against its own reading, marker by marker: were the numbers parsed correctly, was the LOINC mapping right, was the trend read as they would read it, does the source behind each statement hold up. Panels from several dates for the same person and from more than one laboratory are the most telling, because they exercise normalisation and trend reading at once. There is no commitment at that stage; production follows when the team is satisfied, inside its perimeter or a private cloud.

Containers run on your infrastructure or in a private cloud; you own the keys and Vitrubo has no access to the data. Identity never enters the prompt: the models see biomarker values and an internal ID, never a name. Vitrubo works under a HIPAA business associate agreement for US deployments, a GDPR Article 28 processing agreement and Standard Contractual Clauses, and encryption in transit and at rest. Roughly 100,000 anonymised hospital records have been analysed this way, on premises, inside a clinical programme.

Yes, through a clinic, laboratory or health app that offers Vitrubo under its own brand; there is no self-serve signup. The person uploads the PDF or a photo of the result and reads the patient view: which biomarkers stand out, how they connect, and what to bring to the next appointment. Earlier results uploaded alongside are analysed into the same history, so one blood test is read on its own and several are read as a trend. What a good analysis of your own blood looks like is what this page describes: numbers from code, one scale across laboratories, each biomarker in context, and a source behind every statement.