Patient-finding methods
Four ways to find rare disease patients
Every rare disease team eventually asks the same question: how do we find the patients we are missing? There are four common answers, and each has a place. This page compares them on the dimensions that decide whether a program holds up.
Last reviewed:
Side by side
The four methods at a glance.
- Manual chart review.
- A clinician reads individual records, one patient at a time.
- Rule-based EHR alerts.
- A hand-written rule fires inside the EHR when a specific signal appears.
- Black-box predictive models.
- A machine learning model estimates each patient's likelihood of disease.
- Governed computational phenotyping.
- An expert-approved disease definition runs as explicit logic across a whole population's existing data.
| Method | Reproducible? | Explainable? | Scale | Clinician time | Governance & audit |
|---|---|---|---|---|---|
| Manual chart review | Varies by reviewer | High for that chart | A few charts at a time | Hours per patient | Informal |
| Rule-based EHR alerts | High for a single rule | Yes, the rule is visible | Population-wide, one signal at a time | Low per alert; fatigue at volume | Often unowned over time |
| Black-box predictive models | Can shift with retraining | Low: a score without the why | Population-wide | High: scores lack supporting evidence | The model is the definition |
| Governed computational phenotyping | Same definition, same data, same patients | Yes: evidence behind every match | Population-wide | Focused: review a filtered list | Versioned, expert-approved |
Each method is described in detail below, including where the simpler ones are the right choice.
Method 1
Manual chart review
The default, and still the right tool in the right setting. When a specialist already suspects a handful of patients, sitting with the full chart gives context no automated method matches: the note nobody coded, the family history mentioned once. There is nothing to integrate and nothing to buy.
The limits are arithmetic. A careful review takes hours per patient, so nobody chart-reviews a population of two hundred thousand. And the criteria live in the reviewer's head: two clinicians reading the same chart can reach different conclusions, which makes results hard to reproduce or defend at scale.
Method 2
Rule-based EHR alerts
Alerts are inexpensive, live inside the workflow clinicians already use, and work well when one signal is enough: a single lab value, a specific code. For some conditions, that is all you need.
Rare disease rarely presents that way. The signal is usually a combination of labs, symptoms, and history spread across years of encounters. Single-signal rules either miss those patients or fire so often that clinicians stop reading them; alert fatigue is well documented. And in many systems, rules accumulate for years with no owner and no change history, so nobody can say why a given alert still fires.
Method 3
Black-box predictive models
Machine learning models trained to estimate disease likelihood can find patterns nobody thought to specify, and they are a legitimate, active research area. They also run at population scale.
The trade-offs are practical. The output is a probability, and a score alone gives a reviewing clinician little to act on: why this patient? Retraining can change which patients surface, which makes cohorts hard to reproduce between runs. And because the model itself is the definition of who counts, auditing the definition means auditing the model, which few clinical teams are equipped to do.
Method 4
Governed computational phenotyping
Computational phenotyping turns a clinical definition of a disease (the codes, labs, signals, inclusions and exclusions) into explicit logic and runs it across existing data. Governance adds the part that makes it defensible: the definition is curated and signed off by clinical experts, versioned, and auditable. The result is deterministic. The same definition over the same data always surfaces the same patients, each with the evidence that matched.
It has real costs too. The definition takes expert work up front, and the method needs access to structured data such as FHIR. If you have three suspected patients and a free afternoon, chart review is faster. Where this method earns its keep is the population-scale question the others cannot answer reproducibly: who else is out there?
Where PathfindEHR™ fits
An implementation of the fourth method.
PathfindEHR™ applies governed computational phenotyping. Key opinion leaders curate and sign off every disease pathway, the pathway runs across a health system's existing FHIR and real-world data in place, and the output is a filtered, explainable candidate list for a specialist to review. It does not score, rank, or predict patients. It is decision support, not a diagnostic device: the clinician reviews the evidence behind each candidate and makes every decision.