Prognova AI

Intelligence, designed for medicine.

Developing healthcare-specific AI to help doctors work with complex clinical information, grounded in high-quality data and rigorous evaluation.

Explore our technologyOur approach

Behind the headline, a three-dimensional anatomical brain turns slowly. Fine teal pathways run across its surface and short coral pulses travel along them, representing signals moving through a healthcare-specific model.

Clinical information arrives in pieces.

A single patient visit can produce a handwritten note, a laboratory printout with its own reference intervals, an ECG strip, an imaging report, and a prescription in a different format from the last one. Each is correct on its own page. Together they are hard to hold in mind.

General-purpose language models were trained to be fluent, not to be grounded. They do not know which page a value came from, they fill gaps with plausible text, and they are rarely measured against what a clinician actually transcribed from the original report.

Prognova is built around one idea: every piece of information the model works with should stay attached to where it came from, and anything missing, unreadable or conflicting should be reported as exactly that.

How Prognova differs from a general-purpose language model
General-purpose modelPrognova
Trained on general web textTrained on curated healthcare documents with provenance
Fills gaps with plausible textReports missing, unreadable or conflicting values as such
Rarely measured against clinical ground truthMeasured against reviewer-transcribed gold annotations
Privacy handled after the factIdentifier screening before anything is kept

From the printed page to a model you can check.

Three stages, one chain of custody. Data is curated with its provenance intact, a healthcare-specific model is trained only on approved material, and everything is measured against reviewer-transcribed ground truth before it is called ready.

  1. Stage 1: High-quality healthcare data
  2. Stage 2: Healthcare-specific intelligence
  3. Stage 3: Evaluation and clinical validation

Every value keeps its source.

Reports are read page by page. Each extracted value carries the document, the page and the verbatim snippet it came from, and a validator re-reads the source independently to check that the value, label and unit really appear there.

Values are stored exactly as printed: no rounding, no unit conversion. A smudged digit is recorded as unreadable, two numbers on one line as ambiguous, a missing reference interval as not reported. Nothing is guessed.

Before any record is kept, pattern detectors look for direct and indirect identifiers. Records with an unresolved finding are quarantined and accounted for. Every record ends in exactly one outcome: exported, sent for review, rejected, or quarantined.

  • Source-grounded extraction from laboratory reportsAvailable
  • Source-grounded extraction from ECG printoutsAvailable
  • Identifier screening and quarantineAvailable
  • Prescriptions, imaging and discharge summariesResearch

A small model, trained on what it will actually see.

From approved exports we build task datasets: structured extraction, locating a field on a page, question answering that includes unanswerable questions, and template summaries. Splits are made by document so the model is never tested on a page it has seen.

We are selecting a base model against a written framework rather than a leaderboard, and training is gated behind data-quality checks and a human-written approval. No fine-tuning run starts until those gates pass.

The goal is a small language model that is specific to healthcare documents: cheaper to run, easier to evaluate, and honest about what it cannot answer.

  • Task datasets built from approved exportsIn validation
  • Base-model selection frameworkIn validation
  • Healthcare-specific fine-tuningResearch

Measured against what reviewers transcribed, not against itself.

Reviewers transcribe values from the original reports without seeing any pipeline output. Those gold annotations are the only basis for quality figures, and gold documents are held out of training.

Until measured quality meets written release criteria, a dataset is marked development-only. Unsupported and uncertain records are kept and analysed rather than discarded, so error analysis has something to work with.

Clinical testing follows the same discipline: clinician-reviewed annotations carry a reviewer and a date, and the model’s role stays assistive throughout.

  • Gold annotation protocol and release criteriaAvailable
  • Measured extraction qualityIn validation
  • Clinical testing with practising doctorsResearch
Available
Built and in use by the team today.
In validation
Built, being measured against gold annotations before release.
Research
Under investigation; no result to report yet.

The doctor decides. The model keeps the record straight.

Prognova is designed to sit beside a clinician while they read a complex record: surfacing what is in the documents, where it came from, and what is missing. It does not diagnose, and nothing enters the chart without the clinician’s action.

ResearchIllustration of intended use with a fictional case. Not patient data, not a shipped product.

  1. Step 1: Notes
  2. Step 2: Labs
  3. Step 3: Imaging
  4. Step 4: Considerations
  5. Step 5: Follow-up

Clinician

Decides what enters the record

Consultation note

  • Adult patient, fatigue for three weeks, increased thirst and nocturia.
  • Known hypertension, six years. No new medications reported.
AcceptEditDismissClinician confirms the history

Prognova

Surfaces and cites. Never decides.

Reading the record

  • Two prior visits mention fatigue; the earliest is eleven months ago.

    Source: Clinical notes, visit 2 and visit 5

  • Blood pressure medication unchanged since the last prescription on file.

    Source: Prescription, page 1

Clinician

Decides what enters the record

Laboratory report

  • HbA1c 8.2 % (ref < 5.7)
  • Fasting glucose 164 mg/dL (ref 70 to 99)
  • Creatinine 1.1 mg/dL (ref 0.7 to 1.3)
  • Potassium 4.6, unit not printed
AcceptEditDismissClinician accepts the flags and edits one note

Prognova

Surfaces and cites. Never decides.

Values outside the printed interval

  • HbA1c and fasting glucose are above their printed reference intervals.

    Source: Laboratory report, page 2

  • Potassium has no printed unit; flagged for review rather than interpreted.

    Source: Laboratory report, page 2

Clinician

Decides what enters the record

Imaging summary

  • Chest radiograph, two views. Printed interpretation: no acute cardiopulmonary findings.
AcceptEditDismissClinician reads the original report

Prognova

Surfaces and cites. Never decides.

As printed, unattributed

  • The printed interpretation is shown verbatim and is not presented as clinician-confirmed.

    Source: Imaging report, page 1

Clinician

Decides what enters the record

Clinician’s working list

  • The clinician writes the differential. The model does not propose diagnoses.
AcceptEditDismissClinician decides what to order

Prognova

Surfaces and cites. Never decides.

Open questions from the record

  • No lipid panel or urine albumin result is on file in the last twelve months.

    Source: Laboratory reports, all pages

  • Family history section of the intake form is blank.

    Source: Intake form, page 2

Clinician

Decides what enters the record

Plan

  • Repeat laboratory panel in three months. Review home readings at the next visit.
AcceptEditDismissClinician edits, signs, and owns the plan

Prognova

Surfaces and cites. Never decides.

Draft for the clinician to edit

  • A follow-up summary is drafted from the accepted items only, with each line keeping its source.

    Source: Accepted items, this visit

Prognova AI is an assistive technology under development. It is not a substitute for clinical judgement and does not diagnose, treat, or make decisions about patients.

Trust is a process, written down.

What follows is how we work, not what we have achieved. Figures appear on this site only once they have been measured against reviewer-transcribed ground truth.

Data quality

Values stay verbatim and attached to their page.

Every extracted value is stored as displayed, with the document, page and snippet it came from. An independent validator re-reads the source and rejects evidence that cites another document, a wrong page, a blank scan or an altered unit.

Missing, unreadable, ambiguous and conflicting values are recorded as such. Conflicts are flagged, never resolved by the software.

Privacy

Identifiers are screened before anything is kept.

Pattern detectors look for direct and indirect identifiers in every record. Anything with an unresolved finding is quarantined, and the quarantine is counted in the batch report so no record disappears silently.

Processing runs locally with network access blocked. Logs and reports contain counts only, never document text or file paths. We treat every output, including approved datasets, as sensitive health data.

Evaluation

Ground truth comes from reviewers, not from the model.

Reviewers transcribe values from the original reports without seeing any pipeline output. Those gold annotations are the only basis for quality figures, and documents used as gold are held out of training.

A dataset stays marked development-only until measured quality meets written release criteria.

Error analysis

Rejected and uncertain records are kept, not discarded.

Each record ends in exactly one outcome: rejected, review, privacy quarantine, origin excluded, or exported. The review and rejected files are inputs to the next iteration, so every failure mode has a paper trail.

Recall is deliberately low. An unrecognised label or a value without a printed unit is skipped or sent for review rather than guessed.

Human oversight

People approve data, training runs and releases.

Model training needs a human-written approval file and passing quality gates before it can start. Sending data to any external service needs a written, per-batch authorisation, and no such integration exists today.

In the clinic, the model’s role is to surface and cite. Nothing enters a record without a clinician’s action.

Clinical validation

Clinician-reviewed annotations carry a reviewer and a date.

Clinical testing with practising doctors is planned and has not started. When it does, clinician-reviewed annotations will be recorded with the reviewer and date, and results will be published only after review.

Prognova AI holds no regulatory clearance and makes no claim of clinical performance.

Research in the open, when it is ready.

We publish results after they have been reviewed. Until then, these are the questions we are working on and where each one stands.

Publications

No publications yet. Reviewed results will appear here first.

Open questions

  1. Source-grounded extraction from laboratory and ECG reports

    How reliably can values be extracted with page-level provenance, measured against independently transcribed gold annotations?

    In validation

  2. Knowing when the record does not say

    Can a small healthcare model be trained to answer that a question is unanswerable from the documents, and how should that be scored?

    Research

  3. Choosing a base model for a healthcare SLM

    Which open small models are suitable for fine-tuning on healthcare documents, judged by a written framework rather than a leaderboard?

    In validation

  4. Preventing leakage between training and evaluation

    How should a corpus be split and audited so that no page, template or patient-level duplicate crosses from training into test?

    Research

Built by people who think a medical model should show its sources.

Prognova AI is developing a healthcare-specific small language model to help doctors work with complex clinical information. We believe that model should be trained on carefully curated and labelled healthcare data, measured against ground truth transcribed by reviewers, tested with clinicians before it is trusted, and always kept under human oversight.

Grounded
Every output can be traced to a document and a page.
Honest about gaps
Missing, unreadable and conflicting information is reported, not smoothed over.
Assistive by design
The clinician decides. The model surfaces, cites and drafts.
Measured before claimed
No figure appears before it has been measured against gold annotations.

Team

Team profiles will be published here once confirmed by the company.

Talk to us.

We are looking for clinicians, hospitals and researchers who want to help shape and test a healthcare model that shows its sources.

Please do not include health information about yourself or anyone else.

We are setting up our contact inbox. This page will update when it is ready.

No message can be sent from this page yet, and this site stores nothing you type.