THE GOLD STANDARD OF GENETICS · SPECIALIST STUDY
Treat a polygenic score as a versioned statistical model for a defined population and outcome, not as a genetic diagnosis or a destiny label
A polygenic risk score combines information from many genetic variants into one statistical estimate. The number can rank people within a study population, but rank is not the same as probability, diagnosis or clinical benefit. Performance depends on the discovery data, linkage-disequilibrium structure, allele frequencies, genotyping and imputation quality, selected variants, weights, ancestry composition, outcome definition and evaluation cohort. A score that separates cases from controls in one population can be miscalibrated in another: people may be placed too high or too low, and the same percentile may correspond to different absolute risks. Responsible use therefore identifies the exact score and version, verifies technical compatibility, describes ancestry and relevant social context without treating ancestry as a simple biological box, tests discrimination and calibration in a representative external cohort, combines the score with age, family history and conventional risk factors, and reports a time-bounded absolute risk only when an appropriate baseline model exists. A threshold earns clinical meaning only if acting on it improves decisions and outcomes enough to justify cost, anxiety, privacy loss and unequal error. Most scores remain probabilistic research or decision-support tools, not standalone screening diagnoses.
This guide belongs to the Genetics subject root, the Gold Standard of Education, and the How X Works hub. It begins with a concrete decision, then tests whether the method still holds when the context changes.
Choose a reading route
Identify the score and purpose · chapters 1, 2, 3
Check data and portability · chapters 4, 5
Validate predictions · chapters 6, 7, 8
Decide clinical use · chapters 9, 10, 11
Govern and transfer · chapters 12, 13
The whole problem at a glance
| Question | Evidence needed | Unsafe shortcut |
|---|---|---|
| Which score is this? | Version, variants, weights, build and trait | A label identifies a stable product |
| Does it transport? | Target-population external validation | Good discovery AUC applies everywhere |
| What is the probability? | Calibrated baseline and time horizon | Percentile equals absolute risk |
| Should anyone act? | Utility, harms, equity and pathway | High score automatically needs treatment |
Reclassification can occur when the score, genotype pipeline or baseline model changes.
CHAPTER 1 OF 13 · Identify the score and purpose
1. Define the outcome, population and decision
Specify the disease or trait, event definition, age range, prediction horizon, intended population and whether the score is for research, stratification or a particular clinical decision. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A coronary score is evaluated for ten-year risk in adults without prior disease; it is not silently reused to predict lifetime risk in adolescents. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: outcome definition; exclusions; age; time horizon; setting; target population; intended use; decision; alternatives Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: A broad disease label can hide different phenotypes and time horizons, while a useful association does not create an action. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Keep the score fixed but change the horizon from ten years to lifetime and rebuild the baseline model. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 1 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 2 OF 13 · Identify the score and purpose
2. Identify the exact score and its provenance
Record the published score, registry identifier, genome build, variant list, effect allele, weights, training data, selection method and version. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: Two scores share a disease name but use different variants and weights; the laboratory reports the exact PGS identifier instead of treating them as interchangeable. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: PGS ID; citation; version; build; variants; alleles; weights; units; training cohort; licence; update date Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: A score name is not enough for reproducibility, and newer is not automatically better for every target population. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Swap one version for another and require a full technical and clinical bridge. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 2 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 3 OF 13 · Identify the score and purpose
3. Separate ancestry, population and social context
Describe genetic ancestry estimation, self-identified race or ethnicity, geography and environment as distinct variables, using each only for a justified purpose. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A participant has mixed genetic ancestry and lives in a different risk environment from the training cohort; the report avoids forcing a single category and states the portability evidence. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: ancestry method; reference panel; uncertainty; self-identification; recruitment geography; environment; care access; missingness Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: Ancestry is continuous and model-dependent, while social categories can capture exposures and inequity that genetic principal components do not. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Keep genetic ancestry constant but change the environmental baseline and show why absolute risk changes. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 3 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 4 OF 13 · Check data and portability
4. Check genotyping, sequencing and harmonisation
Confirm sample identity, variant call quality, genome build, strand, effect allele, imputation, dosage handling and coverage of required variants before calculating the score. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A palindromic variant with ambiguous strand is removed under the score protocol rather than assigned an effect allele by guesswork. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: sample ID; platform; call rate; imputation panel; INFO score; build; lift-over; strand; allele frequency; missing-variant rule Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: A numerical output can be produced from mismatched alleles or builds and still look plausible, so reproducibility starts before statistics. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Reverse one effect allele and quantify how the score and percentile change. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 4 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 5 OF 13 · Check data and portability
5. Handle missing variants and pipeline drift
Follow pre-specified rules for missing, poorly imputed or ancestry-divergent variants and validate any proxy substitution or rescaling. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A platform lacks three weighted variants; the laboratory reports reduced completeness and does not renormalise unless the score’s validated method permits it. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: variant completeness; missing weights; proxies; LD reference; rescaling rule; software version; container; checksums; deviations Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: Unreported renormalisation can change meaning, and a proxy effective in one ancestry may be weak in another. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Change the LD reference panel and require new validation of each substituted proxy. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 5 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
Worked PRS interpretation
| Observation | Possible meaning | Needed check | Do not infer |
|---|---|---|---|
| 95th percentile | High relative position in reference set | Which reference population and version | 95% chance of disease |
| Hazard ratio 2.0 | Relative association in that model | Baseline risk, age and calibration | Risk is exactly doubled for everyone |
| AUC improves 0.02 | Small discrimination change | Calibration and net clinical benefit | The test improves outcomes |
| Lower performance in one ancestry group | Portability or data imbalance | Representation, QC and recalibration | Ancestry is a fixed genetic essence |
Each layer answers a different question and needs different evidence.
CHAPTER 6 OF 13 · Validate predictions
6. Measure ancestry portability directly
Evaluate the complete score in external target cohorts, comparing distributions, effect size, discrimination and calibration across relevant ancestry and mixed-ancestry groups. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A score developed mainly in European-ancestry data ranks an East Asian cohort differently; local validation reveals a shifted mean and overestimated tail risk. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: external cohort; recruitment; ancestry distribution; sample size; case definition; effect estimate; AUC; calibration; uncertainty; subgroup results Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: A high development-set result does not guarantee transport, and subtracting an ancestry-group mean does not prove equal performance. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Keep AUC stable but shift calibration and show why individual probabilities remain wrong. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 6 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 7 OF 13 · Validate predictions
7. Separate discrimination from calibration
Report how well the model ranks higher- and lower-risk people and whether predicted probabilities match observed frequencies, with confidence intervals and calibration plots. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: Two models have the same AUC; one predicts 20 percent risk for a group with 10 percent observed events, so it is not clinically interchangeable. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: AUC or C-index; sensitivity; specificity; calibration intercept; slope; observed-to-expected; plots; uncertainty Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: AUC can remain unchanged after systematic overprediction, and a statistically significant association says little about probability accuracy. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Apply a constant upward shift to every prediction and compare discrimination with calibration. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 7 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 8 OF 13 · Validate predictions
8. Translate relative position into absolute risk carefully
Combine the score with an appropriate baseline incidence or validated clinical model, stating age, sex where relevant, competing risks, time horizon and population. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A person in the 90th percentile receives a five-year absolute-risk range based on local incidence and conventional factors, not a statement that ninety percent will develop disease. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: reference distribution; percentile; effect scale; baseline incidence; age; competing risks; clinical covariates; uncertainty; model version Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: Percentile, odds ratio and absolute probability are different quantities, while rare outcomes can remain uncommon at high relative risk. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Keep the percentile fixed but lower baseline incidence and recalculate the absolute risk. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 8 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 9 OF 13 · Decide clinical use
9. Test incremental and clinical utility
Compare decisions with and without the score using pre-specified thresholds, reclassification, decision-curve or trial evidence, downstream adherence and meaningful outcomes. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: Adding PRS changes a few preventive-treatment decisions, but benefit is claimed only after net benefit and harms are evaluated. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: comparator model; threshold; reclassification; net benefit; trial or implementation evidence; outcomes; adherence; cost; harms Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: A small change in AUC may or may not improve care, and better prediction does not guarantee better outcomes. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Hold model metrics constant but remove an effective action and explain why utility disappears. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 9 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
Polygenic-score gate
| Gate | Pass evidence | Remain open when |
|---|---|---|
| Identity | Score and pipeline are reproducible | Version or genome build is unclear |
| Portability | Representative external validation exists | Only discovery data are reported |
| Calibration | Predicted and observed risks agree | Only AUC or odds ratios are shown |
| Utility | An action pathway improves decisions fairly | Thresholds are chosen without outcomes |
A score may be scientifically informative yet not ready for individual care.
CHAPTER 10 OF 13 · Decide clinical use
10. Integrate monogenic, family and conventional risk
Keep pathogenic high-penetrance variants, family history, age, biomarkers, lifestyle and existing clinical scores visible rather than letting PRS absorb or duplicate them. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A patient with familial hypercholesterolaemia is managed through the monogenic pathway even if the polygenic score is average. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: monogenic testing; pedigree; clinical measurements; exposures; medicines; conventional model; dependencies; conflicts Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: An average PRS does not cancel a pathogenic variant or strong phenotype, and family history contains genetic and non-genetic information. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Add a confirmed monogenic variant and rebuild the hierarchy of evidence. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 10 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 11 OF 13 · Decide clinical use
11. Set thresholds around an explicit action
Choose any high-, intermediate- or low-risk threshold by the action it triggers, the evidence for benefit, resource capacity and acceptable false-positive and false-negative trade-offs. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A screening threshold is tested against programme capacity and missed cases instead of being set at an attractive round percentile. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: action; threshold; sensitivity; specificity; PPV; NPV; capacity; benefit; harms; equity; exceptions Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: A tail percentile is not a natural clinical boundary, and thresholds can create unequal error across populations. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Move the threshold from top five to top ten percent and calculate who gains, who is harmed and whether capacity holds. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 11 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 12 OF 13 · Govern and transfer
12. Communicate probability without genetic determinism
Explain relative and absolute risk, uncertainty, modifiable and non-modifiable factors, non-diagnostic status and what the result does or does not change today. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A high score is described as one risk factor within a prevention plan, while an average score is not used to excuse smoking or ignore symptoms. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: result; reference group; absolute range; uncertainty; action; family implications; psychosocial response; teach-back; questions Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: Genetic information can feel fixed, but a probabilistic score neither guarantees disease nor guarantees protection. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Keep the number fixed but change the available preventive action and rewrite counselling. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 12 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
CHAPTER 13 OF 13 · Govern and transfer
13. Govern equity, privacy and reanalysis
Audit performance and uptake by subgroup, protect genomic data, limit secondary use, document version changes and define when people will be recontacted. Begin with one versioned polygenic score, one defined outcome and time horizon, one person and ancestry context, one technical pipeline, one target population, one proposed action and one review date. State the decision, specimen, record or person, the time boundary and the responsible owner before interpreting any result.
Worked example: A health system detects lower calibration and lower follow-up access in one community; it pauses expansion and corrects both model and service barriers. Put the starting evidence, applicable rule, intermediate decision, later observation and consequence on separate lines so that a fluent story cannot replace a reproducible pathway.
Evidence to retain: consent; data access; retention; secondary use; subgroup audit; access; complaints; score version; recontact; withdrawal Preserve provenance, dates, versions, identity checks, responsible people and intended use. If a necessary item is missing, label the gap and its effect instead of converting absence into reassurance.
Boundary and misconception: Equal availability is not equal benefit when performance and downstream care differ, while silent model updates can change classifications. Test the strongest credible alternative explanation and name the new evidence that would change the conclusion. The pass condition is a bounded, reviewable claim, not the removal of uncertainty.
Evidence discipline: record the observation, version or time boundary that supports this chapter’s conclusion. Preserve the rejected alternative as well as the selected action. That makes a later review able to distinguish a changed world from an undocumented change of mind.
Changed-context transfer: Improve model calibration but remove follow-up access and show why the programme still fails. The method transfers only if the decision question and measurement meaning survive the change. If either changes, rebuild the comparison rather than copying the earlier threshold or weight.
Chapter 13 check · what should survive an audit
- Name the exact decision, unit, time boundary and responsible authority.
- Preserve the supporting evidence, controls and strongest credible alternative.
- Record the next action, review date, stop rule and reopening condition.
Verification lab: work the decision twice
Take a fictional PRS at the 95th percentile and write three separate statements: relative position, validated association and time-bounded absolute risk.
A score has good AUC but overpredicts disease in the target population. Design a recalibration and external-validation plan before any clinical threshold is used.
A mixed-ancestry patient falls outside the score’s main evaluation groups. Build a transparent uncertainty statement and preserve conventional risk assessment.
Final review: distinguish score identity, technical validity, association, discrimination, calibration, absolute risk and utility. If any two are treated as synonyms, the pathway fails.
Limits and misconceptions
A percentile is not a probability
It describes relative position in a reference distribution, not the chance of disease.
Portability must be measured
Ancestry and population differences can change distributions, effect estimates and calibration.
Prediction is not utility
A model can predict better without improving decisions or outcomes.
Average is not protection
An average or low PRS does not cancel symptoms, monogenic variants, family history or conventional risk.
Questions readers ask
What is a polygenic risk score?
- A weighted sum of many genetic variants associated with a defined trait or disease.
Does a high PRS mean I will develop the disease?
- No. It changes probability within a population and context; it is not a diagnosis or certainty.
Why does ancestry matter?
- Variant frequencies, linkage patterns and training representation affect score distributions and performance.
Is AUC enough to validate a score?
- No. Calibration, external validation, absolute risk and clinical utility are also needed.
Can PRS replace family history?
- No. Family history captures monogenic, polygenic, environmental and social information not contained in one score.
When is a PRS clinically useful?
- When a validated, calibrated result changes a proven action in a way that improves net benefit fairly.
Practise, verify and continue
Build a PRS record with score ID, version, variants, build, pipeline, target trait, horizon, development and validation populations, ancestry method, completeness, discrimination, calibration, baseline model, absolute-risk calculation, proposed threshold, action, utility evidence, equity audit and update trigger.
Apply the method to an AI clinical risk model. Preserve external validation, calibration, absolute risk and utility, then rebuild data provenance and drift controls because inputs are not fixed germline variants.
Practical evidence check
Before implementation, write a one-page evidence map with the decision, unit of analysis, version or time boundary, inputs, outcomes, uncertainty, authority and stop rule. Attach one worked case that passes and one that fails. A reviewer should be able to reproduce the reasoning without relying on the memory of the person who built the system.
Run one changed-context rehearsal before approval. Alter a mechanism rather than a cosmetic detail, then identify which calculation, safeguard or governance gate must change. Record any conclusion that does not transfer. This short rehearsal is often the fastest way to expose a hidden assumption before it becomes an operational dependency.
Related eduKate reading
Genome sequencing, panels and coverage gaps
Carrier screening, ancestry and residual risk
Variant classification and evidence weight
Pharmacogenomics and clinical decision support
Authoritative sources and scope
National Human Genome Research Institute: Polygenic Risk Score
NHGRI: reporting standards for polygenic risk scores
PGS Catalog: score and performance data documentation
PGS Catalog: ancestry representation framework
These sources support the laboratory, statistical and counselling distinctions described here. They do not replace patient-specific evaluation, informed consent, validated laboratory methods, local regulation or advice from a qualified genetics professional. All worked cases are fictional. A result must be interpreted with phenotype, family history, ancestry context, assay limits and current evidence; it must not be used alone to diagnose or change care.