terrals

01 · Large-scale multi-modal

Every layer of the patient, linked.

Clinical records, radiology, whole-slide pathology and omics for the same people — joined on one de-identified key and ordered on one timeline, so your model learns medicine the way it actually unfolds.

The Terrals record

One patient. Every layer.

Most datasets give you one slice of a person. We link every signal the hospital captured, in the order it happened.

  1. 01Clinical timeline
  2. 02Radiology
  3. 03Pathology
  4. 04Omics
  5. 05Linked & de-identified

Diagnoses, labs, medications, procedures and notes — years of encounters on one axis.

2019202020212022202320242025EMRImagingH&EOmicsHypertensionLipid panelBreast lumpCore biopsyChemoSurgeryFollow-upWSI · 40×core biopsyPIK3CATP53RNA-seq · tumour vs normalpatient · tr_8f3a…c21 · de-identified
Illustrative, synthetic patient. No real patient data is shown on this site.

The training stack

From the first token to the last reward.

Whatever stage your model is in, we supply the clinical signal it needs next.

01Pre-training

Corpus-scale clinical data.

De-identified notes, reports, DICOM, whole-slide images and sequence data — deduplicated, tokenisation-ready and linked across modalities.

  • Clinical text
  • Image patches
  • Sequence k-mers
  • Multi-modal pairs
dyspnoeaonexertionHbA1c8.1%→metforminECOG1biopsyIDCgradeIIdyspnoeaonexertionHbA1c8.1%→metforminECOG1biopsyIDCgradeII
ER+PR+HER2-Ki-6730%mitosesnuclearpleomorphismstromaER+PR+HER2-Ki-6730%mitosesnuclearpleomorphismstroma
ACGTTGGATTCATTAGGCCCGATAAGCTTAGGCATCTACGGACTTAAGACGTTGGATTCATTAGGCCCGATAAGCTTAGGCATCTACGGACTTAAG
your model
02Fine-tuning

Domain adaptation, by specialty.

Curated cohorts for oncology, cardiology, radiology and more — balanced, labelled and split for continued training.

  • Specialty cohorts
  • Report–image pairs
  • Labels & outcomes
general+ specialty datasteps →eval loss
03SFT

Expert-written demonstrations.

Instruction–response pairs authored and reviewed by practising clinicians, grounded in real (de-identified) cases.

  • Clinician-authored
  • Case-grounded
  • Multi-turn
62F, breast lump 2.1 cm, core biopsy: IDC grade II, ER+/PR+/HER2−. What's the next step?
authored by oncologist · reviewed ×2
04RLHF

Clinician preference at scale.

Panels of doctors compare and rank model outputs with rubric-based rationale — the signal reward models need.

  • Pairwise ranking
  • Rubrics
  • Rationales
Response A

Start chemotherapy immediately for all patients.

  • ✗ Safe
  • ✗ Guideline-concordant
  • ✓ Cites evidence
Response B

Surgery first; use genomic risk to decide on chemo.

  • ✓ Safe
  • ✓ Guideline-concordant
  • ✓ Cites evidence
reward model signal…
05RL environments

Rewards from real outcomes.

Environments where the agent's plan is scored against what actually happened next in the patient's record.

  • Outcome-anchored
  • Leakage-checked
  • Deterministic reward
Agentepisode ↻Staterecord to day 0Actionorders · meds · planRewardvs. real day 90

How an engagement works.

You tell us

  • Disease area or cohort definition
  • Modalities and how they must be linked
  • Inclusion & exclusion criteria
  • Follow-up window
  • Volume and delivery cadence

You get

  • One de-identified patient key across every modality
  • Event-level clinical timeline (diagnoses, labs, meds, procedures, notes)
  • Native DICOM and whole-slide images with their reports
  • Data dictionary and cohort summary
  • Encrypted, access-controlled delivery

Process

  1. 1

    Scope

    We agree the cohort, modalities and linkage you need.

  2. 2

    Feasibility

    We check what exists before you commit, and share the counts.

  3. 3

    Sample

    A small de-identified sample so your team can check format and quality.

  4. 4

    Delivery

    The full cohort, documented and delivered securely.

Questions.

How is a patient linked across modalities?+

Inside the hospital, every record is mapped to one study key before identifiers are removed. Only that key leaves; the mapping does not.

Can we see data before committing?+

Yes. We share a feasibility summary first, then a small de-identified sample for you to evaluate.

What formats do you deliver?+

Tabular clinical data as Parquet or CSV, imaging as native DICOM, pathology as SVS/TIFF, sequencing as FASTQ/BAM/VCF — or to your schema.

Where does the data come from?+

Our data comes from partner hospitals and research institutes we work with directly.

Request data

What does your model need?

Tell us the modalities, scale and timeline. We'll come back with a sample and a scoped proposal.

Prefer email? hello@terrals.bio

What data do you need?

Explore more