Explainer · Real-world evidence
Most of what we know about a patient is buried in free text. How natural language processing gets it out, and why health economists should care.
Assumptions Aside4 min read
In health care, the most important facts are often not in the fields a database was built to hold. They sit in pathology reports, clinic notes and discharge summaries, written by clinicians for other clinicians. Natural language processing, the branch of AI that lets computers read human language, is how that knowledge gets out of the text and into analysis.
The four moves
Clinical NLP comes down to four steps, each building on the last.
- 1
Concept extraction
Spot the clinically meaningful terms in the text: findings, procedures, drug names. Technically this is named-entity recognition.
- 2
Normalization
Map different spellings, abbreviations and older names to one concept, so that “fibrotic” and “fibrosing” are counted as the same thing.
- 3
Context detection
Decide whether a mention actually counts. Is it negated (“no evidence of”)? Hedged (“cannot exclude”)? Historical or current? A lab value, or a diagnosis?
- 4
Phenotyping
Roll the findings up into a patient-level label, such as “likely has the disease”, that analysts and models can use.
Try it: step through how a lexicon reads a synthetic note.
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
Rules, learning, or both
There are two broad approaches. Rule-based NLP is driven by a curated dictionary and context rules. It is transparent and auditable: you can see exactly why a patient was flagged. Machine-learning NLP learns patterns from labeled examples. Strong projects usually combine them, with a lexicon and rules first and machine learning layered on top.
The lexicon is upstream of everything. If the dictionary is wrong, every model built on it is wrong too.
Where the lexicon sits
A phenotyping lexicon is the evidence-backed dictionary that tells the software what to look for, which variants count, how strong each term is as evidence, and when to ignore a match. It is the first stage of a five-stage pipeline.
- 01
Build & verify the lexicon
The evidence-backed dictionary of terms, variants, strengths and exclusion rules.
- 02
Extract
Software reads reports, notes and medication records, applies context rules, and records findings per patient.
- 03
Validate
Clinicians hand-review a sample of charts to measure precision and recall. Terms and rules are tuned.
- 04
Model
Machine learning combines the extracted findings with structured data, such as lab values and diagnosis codes.
- 05
Apply
The validated model runs across the full patient population.
Two numbers decide whether the pipeline works. Precision is how often a flagged match is right. Recall is how many of the true cases are found. Clinician chart review measures both, and the lexicon is tuned until they are good enough to trust.
Pathology reports are the strongest evidence. But many patients are diagnosed without a biopsy. Reading clinic notes and treatment history, alongside structured data, keeps those patients from being missed.
Why health economists should care
Finding the invisible patients
Diagnosis codes alone undercount rare diseases. Text-based phenotyping surfaces suspected and undiagnosed cases.
Better prevalence and burden estimates
These are the inputs that every economic model, budget-impact analysis and HTA submission depends on.
Credible real-world evidence
Cohorts for treatment-pattern and disease-progression studies: the evidence payers and HTA bodies increasingly ask for.
Garbage in, garbage out
The lexicon is upstream of everything. A rigorous, source-verified dictionary is what makes the downstream AI trustworthy and defensible.
See how we applied this to a rare-disease case-finding project.
Read the engagementWork with us
We design and verify phenotyping lexicons and the evidence standards that sit around them, so the AI downstream can be defended in front of a payer or an HTA body.