Worked example · Benchmarking method
Anatomy of a diagnosis-gap model.
A walk-through of the Bailiwick method on a hypothetical rare genetic condition: ultra-rare, usually de novo, with a gene–disease link discovered about a decade ago. The numbers are illustrative. The reasoning, the structure and the kinds of findings it produces are the real thing.
True prevalence, triangulated from independent lines of evidence, with an honest range.
How many have a molecular diagnosis today, by age band, in unique patients.
Which levers add the most diagnosed patients per unit of effort, and which barely matter.
Step 1 · Triangulate
No single source is trusted. Several are compared.
For most rare genetic conditions there is no population-based incidence study. So birth prevalence is estimated several independent ways, each with different weaknesses, and the results are laid side by side.
- Cohort share: how often the gene turns up in large sequencing cohorts of the presenting phenotype, scaled by how common that phenotype is.
- Case accrual vs an analog: how fast cases are being reported compared with a better-characterized condition whose prevalence is known.
- Variant-level cross-check: recurrent variants, population frequencies and mutation rates, used to test plausibility rather than as an anchor.
- Ceilings: the maximum the biology allows, first for all de novo variants and then for only those of a pathogenic class.
Where the methods agree, confidence rises. Where they don’t, the range says so and the model carries it forward.
Birth prevalence by method
Illustrative. Per 100,000 live births, log scale. Dot = central estimate; line = range; shaded band = consensus range.
Show as table
| Method | Central | Low | High | Role |
|---|---|---|---|---|
| A Cohort share | 0.12 | 0.08 | 0.16 | Floor |
| B Case accrual vs analog | 0.30 | 0.18 | 0.50 | Central anchor |
| C Variant-level cross-check | 0.10 | 0.02 | 0.10 | Cross-check |
| D′ Ceiling, pathogenic class | 0.60 | – | – | Adjusted ceiling |
| D Ceiling, all de novo | 3.50 | – | – | Absolute ceiling |
| Consensus (geometric mean of A and B) | 0.19 | 0.08 | 0.60 |
Step 2 · Follow the path
Patients don’t arrive labelled with a gene.
They arrive with a symptom: seizures, low muscle tone, a growth problem, a developmental delay. The funnel starts there and follows a birth cohort through each step to a molecular answer. Every step is a place where a child can be lost, and each kind of loss has a different fix.
The final step matters as much as the first. A test that finds a variant but reports it as uncertain leaves a family found but without an answer.
Per 100 affected children born
Illustrative. Where the cohort is lost, and what that leak is called.
- 100Born with the condition
- 88Present with the hallmark feature · atypical presentations missed
- 74Recognized and referred to a specialist · delayed or wrong referral
- 52Genetically tested · no test ordered
- 46Tested with something that can find the variant · wrong test or panel gap
- 37Reported pathogenic or likely pathogenic: diagnosed · VUS
In a real engagement each conversion rate is sourced from the literature, testing data or claims, and flagged where it is assumed.
Step 3 · Separate true from diagnosed
The undiagnosed are rarely spread evenly.
When a gene–disease link is recent, children born since then are tested with modern tools. Adults were assessed years earlier, if at all, and are rarely re-tested. In this illustration, adults are over half of all patients but fewer than one in ten has an answer. Adjust the assumptions to see how the picture changes.
True vs diagnosed prevalence by age band
Illustrative population of 800 patients, unique patients.
- Diagnosed
- Undiagnosed
Show as table
| Age band | True prevalent | Diagnosed | Dx rate |
|---|
True prevalent = population × birth prevalence × survival to each age band. Diagnosed = true × P(tested since the gene was discovered) × P(test covers the gene) × P(reported P/LP). Illustrative values; no statistical comparison applied.
Scenarios
Step 4 · One lever at a time
The biggest lever is often not the obvious one.
Each scenario changes a single input from the central case, so its effect can be read cleanly. In this illustration, better panels and fewer VUS are worth doing but move the number little, because that isn’t where most patients are lost. The backlog sits with adults who were assessed before the gene was known.
For a sponsor, a finding like this changes where field medical, sponsored testing and EHR prompts should point. For families, it can mean an adult who has gone decades without an explanation finally gets one.
Adult re-testing, 12% → 50%
Re-testing adults in the specialty and residential settings where the undiagnosed backlog sits.
Pediatric first-tier testing to 90%
Guideline-aligned, sequencing-first testing at first presentation.
Gene on every panel
Closes the coverage gap. Useful, but coverage is not the bottleneck here.
VUS rate cut by half
Functional evidence and reclassification. Small in count, but each one turns “found” into an answer, and sometimes into a treatment decision.
Additional diagnosed patients versus the illustrative central case of 122. Unique patients.
The possibilities
Match the lever to the leak.
Every condition loses patients in a different place. Once the leaks are sized, the levers rank themselves.
| Where patients are lost | Levers that address it |
|---|---|
| Never suspected; atypical or adult presentation | Clinician education, EHR-embedded prompts, phenotype-driven case finding |
| Suspected but not referred or tested | Sponsored no-charge testing, referral pathways, first-tier testing guidelines |
| Tested with something that can’t find it | Panel content work with labs, exome/genome-first policies, periodic reanalysis |
| Variant found but reported as VUS | Functional characterization, reclassification and recontact programs, data sharing |
| Assessed before the gene was known | Adult re-testing, reanalysis of prior exomes, registry outreach |
| Relatives at risk, never tested | Cascade and family testing with counseling support |
| Found too late to matter | Newborn screening, including genomic newborn screening nominations |
Step 5 · Say what is assumed
Every deliverable separates what is known from what is assumed.
For a condition with few described patients, the honest answer is a range. The model shows which assumptions drive it and what data would narrow it.
Typically well anchored
- Population denominators: births and age structure from vital statistics and census
- Incidence of the presenting phenotype from prospective population studies
- Phenotype and inheritance from published case series
- Testing rates for the broader phenotype group from published cohorts
Typically assumed, and flagged
- The gene’s share of the phenotype when it rests on a handful of cases
- Gene-specific panel inclusion and VUS rates
- Adult survival where it has never been described
- Adult re-testing rates, which are rarely published and often drive the answer
The next step in a client engagement is to narrow whichever assumption matters most, for example by using claims data to estimate how often adults with the presenting phenotype have actually been tested.
What would this look like for your asset?
A Benchmarking engagement builds this model for your indication from real data, as a live model with every input referenced, in three to five weeks.
All values on this page are illustrative and do not describe any specific condition, product or client.