AI · Real-world evidence
An auditable, source-verified NLP phenotyping lexicon. It lets software flag likely patients with a rare, frequently misdiagnosed immune-mediated disease and tell them apart from the conditions it imitates.
- Client
- DreamBig Health, AI-assisted diagnosis project
- Role
- Lexicon build & source verification
- Period
- 2026
100%
of terms traced to a cited source
3
kinds of clinical text covered
4
downstream pipeline stages it feeds
The challenge
Rare diseases hide in free text. Many patients never receive a specific diagnosis code, so claims and EHR data cannot see them. This disease is also routinely mistaken for look-alike conditions, including cancer. Undercounting at the start weakens every prevalence estimate, economic model and HTA submission built on the data.
What we did
- 01Built a concept dictionary of the language clinicians use in pathology reports, clinic notes and medication records, from diagnostic histology to flares, steroid dependence and rescue therapy.
- 02Mapped variant spellings, abbreviations and historical names to a single concept.
- 03Graded how strong each term is as evidence, and wrote exclusion rules for when a match should be ignored. A lab value on its own is not a diagnosis.
- 04Traced every term to its published source in a verification log, so a reviewer can see exactly why each entry is there.
How a phenotyping lexicon reads a note, in four steps. The note is synthetic, and no part of the client’s lexicon is shown.
Synthetic clinical note
Biopsy shows dense inflammatory infiltrate with fibrosing change. No evidence of malignancy. Serum marker elevated; cannot exclude an alternative diagnosis. Responded well to steroid taper.
Concept extraction
Spot the clinically meaningful terms in free text.
- dense inflammatory infiltratePathology finding: strong evidence
- fibrosing changeVariant of “fibrosis”: same concept
- No evidence of malignancyNegated look-alike: mimic ruled out
- Serum marker elevatedLab value: supports, but not a diagnosis
- cannot excludeHedge: lowers certainty
- Responded well to steroid taperTreatment response: supports
Where the lexicon sits in the pipeline
- 01Build & verify the lexicon
- 02Extract
- 03Validate
- 04Model
- 05Apply
The outcome
- The foundation of an NLP and machine-learning pipeline that runs extract, validate against clinician chart review, model, then apply.
- Lets software separate likely cases from their mimics and grounds real-world prevalence and cohort studies.
- Reaches patients diagnosed without a biopsy, through clinic notes and treatment history rather than pathology alone.
The project is ongoing and internal to the client. It is described here at a high level only.
Next engagement
A strategic plan to close infant mortality gaps
Work with us
Tell us about the launch, the coverage question or the program that needs to show its value. We will tell you what evidence it needs, and what it can do without.