Independent and not affiliated with the FDA, MHRA, ISPE, PDA, or any agency. Get the appgoutham@madhadi.com
madhadi.comData Integrity & GxP Quality
Browse all topics → Articles Templates & Procedures Learning paths GlossaryScenariosToolsRegulatory ReferencesLearning PathsTopics About Start here
Risk Assessment Plug-and-play starting point AI & Automation

Risk Assessment: AI Training Data Bias and Representativeness

A plug-and-play risk assessment for bias and representativeness in AI or machine learning training data: methodology, scoring scales, selection bias, class imbalance, historical bias, measurement bias, subgroup-sliced evaluation, mitigations, and the decision to remediate versus explicitly scope out an under-represented subgroup, with a worked specimen.

Document type: Risk Assessment

Read and copy the template below into your own quality system. It is a generic starting point for your own internal use, provided as is, with no warranty; see the Terms and License. Adopting it does not by itself create compliance.

This is a ready-to-use risk assessment. Replace every <<FILL: ...>> placeholder with your own specifics. A worked filled specimen follows so you can see how a completed assessment reads. Verify each cited regulation against the current source before you rely on it.

Document control header

FieldEntry
Assessment titleAI Training Data Bias and Representativeness Risk Assessment
Document number<<FILL: RA-ID, e.g. RA-AI-027>>
Version<<FILL: version>>
Effective date<<FILL: date>>
Model / dataset assessed<<FILL: model name and dataset version>>
Assessment owner<<FILL: role, e.g. Data Science Lead>>

1. Purpose and scope

This assessment identifies and scores sources of bias in a candidate training dataset that could cause an AI/ML model to perform unevenly across the population it will face in production, and it records the decision to remediate or explicitly scope out any subgroup the data under-represents. It applies to the dataset named above, at the point captured by the dataset version in the header, and is re-run whenever the dataset is materially re-extracted or the model is retrained on new data.

2. Methodology

This assessment scores each identified bias source on likelihood of affecting model behavior and severity of consequence if it does, following a risk-based approach consistent with ICH Q9 (Quality Risk Management). It is not a full FMEA; it is scoped specifically to data bias and representativeness, feeding into the broader AI risk assessment for the model as a whole (see AI risk assessment for GxP).

Likelihood scale

ScoreDefinition
1, LowThe gap is small and the affected population is a routine, well-understood case
2, MediumThe gap is moderate, or affects a case the model will encounter regularly
3, HighThe gap is large, or affects a case central to the model’s intended use

Severity scale

ScoreDefinition
1, LowAn error on the affected subgroup has minimal consequence and would be caught by routine review
2, MediumAn error on the affected subgroup could affect a quality decision but with a compensating control in place
3, HighAn error on the affected subgroup could affect product quality, patient safety, or a release decision with no independent check

Risk score = Likelihood x Severity. Score 1-2 is Low, 3-4 is Medium, 6-9 is High. A High score requires remediation or an explicit, QA-approved scope exclusion before the model proceeds to build or retrain.

3. Bias sources assessed

Assess each of the following categories against the actual dataset; not every category will apply to every model.

  • Selection bias. The data captures only some of the population the model will face: one site, one product, one instrument, one time period, or one operating condition.
  • Class imbalance. The outcome of interest is rare in the data, so naive accuracy is meaningless and the model can learn to default toward the majority class.
  • Historical bias. The data reflects past practice, including past inconsistent categorization or past errors, which the model will reproduce faithfully rather than correct.
  • Measurement bias. Different instruments, methods, or analysts produce systematically different readings that the model mistakes for a real difference in the underlying process or population.
  • Label bias. Ground-truth labels were assigned inconsistently across labelers, time periods, or sites, so the “truth” the model learns is itself unreliable in places.

4. Assessment table

#Bias sourceCategoryDescription of the gap foundLikelihoodSeverityRisk score / levelMitigationResidual risk
1<<FILL>><<FILL: selection / imbalance / historical / measurement / label>><<FILL>><<FILL: 1-3>><<FILL: 1-3>><<FILL>><<FILL: remediate or scope out>><<FILL>>
2<<FILL>><<FILL>><<FILL>><<FILL>><<FILL>><<FILL>><<FILL>><<FILL>>
3<<FILL>><<FILL>><<FILL>><<FILL>><<FILL>><<FILL>><<FILL>><<FILL>>

Add rows as needed. Every High-risk row must resolve to either a completed mitigation with re-measured residual risk, or a documented, QA-approved scope exclusion; neither may be left open at sign-off.

5. Subgroup-sliced evaluation requirement

Aggregate performance metrics hide subgroup failure. For every subgroup identified as a bias source above with Medium or High risk, the model’s performance must be reported sliced by that subgroup, not only in aggregate, as part of the model’s validation evidence (not this assessment). Record here which subgroups require sliced reporting and the metric to be used for each.

Subgroup requiring sliced evaluationMetricMinimum acceptable performanceReference
<<FILL>><<FILL: e.g. recall on critical class>><<FILL>><<FILL: validation protocol ID>>

6. Remediate versus scope-out decisions

For each subgroup where full remediation (collecting more representative data) is not immediately feasible, document the explicit decision to exclude it from the intended use rather than deploy the model against it silently.

SubgroupRemediate or scope outRationaleWhere the exclusion is stated in the intended-use documentation
<<FILL>><<FILL>><<FILL>><<FILL: reference to intended-use statement or SOP>>

7. Acceptance criteria

  • Every applicable bias category (selection, imbalance, historical, measurement, label) has been assessed, even where the finding is “not applicable, with rationale.”
  • Every identified bias source is scored for likelihood and severity using the scales in section 2.
  • Every High-risk finding has either a completed mitigation with re-measured residual risk, or a documented, QA-approved scope exclusion.
  • Subgroups requiring sliced evaluation are named with a defined metric and threshold, feeding the model validation plan.
  • The assessment is re-run when the dataset is materially re-extracted or the model is retrained on new data.

8. References

ICH Q9, Quality Risk Management, for the risk-based methodology this assessment follows. FDA, Data Integrity and Compliance With Drug CGMP, Questions and Answers (Guidance for Industry, December 2018), for the completeness and representativeness expectations underlying section 3. GAMP 5 Second Edition (ISPE, 2022), for the risk-based lifecycle context (reference by title, do not reproduce clauses).

Confirm the current version and clause numbers of each reference before issue.

9. Revision history

VersionDateAuthorSummary of change
<<FILL: 1.0>><<FILL: date>><<FILL: author>>Initial issue.

10. Approvals

RoleNameSignatureDate
Author (Data Science)<<FILL>>
Reviewer (Process SME)<<FILL>>
Approver (Quality Assurance)<<FILL>>

Filled specimen

The following shows the assessment completed for an illustrative automated visual-inspection model that screens injectable vials for cosmetic and critical defects. Numbers and findings are illustrative; replace them with your own.

#Bias sourceCategoryDescription of the gap foundLikelihoodSeverityRisk score / levelMitigationResidual risk
1Line 2 under-representedSelectionTraining images 88% from Line 1, 12% from Line 2; Line 2 has a different camera angle and lighting rig339, HighCollected 4,000 additional Line 2 images over 6 weeks; re-balanced to 55/452, Medium, re-assess after next production quarter
2Critical defect class rarityClass imbalanceCritical (fill-level) defects are 0.4% of labeled images339, HighMetrics shifted from accuracy to recall on the critical class, with a fixed minimum recall gate; targeted oversampling of critical-defect images during training4, Medium, recall gate carried into validation protocol
3Labeler drift over 2 yearsLabel biasLabeling guidance for “cosmetic, non-rejectable” tightened in year two; older labels may be inconsistent with the current standard224, MediumRe-labeled a stratified sample of year-one images under the current guidance; agreement with original labels 81%, acceptable for cosmetic-tier only2, Low

Subgroup-sliced evaluation requirement: Line 2 and the critical-defect class both require sliced reporting in the validation protocol, with a minimum recall of 98% on the critical class specifically, not just in aggregate.

Remediate versus scope-out: Line 2 was remediated (data collected) rather than scoped out, because Line 2 represents a third of production volume and excluding it from intended use would leave a material gap in inspection coverage. This is recorded in the model’s intended-use statement, IU-VISION-2026-04, as fully in scope following remediation.

Common inspection findings this risk assessment prevents

  • A model is deployed with no documented assessment of whether its training data represents the population it inspects or predicts against.
  • A known under-representation is discovered only after a production failure, rather than assessed and disposed of before build.
  • Class imbalance is present but performance is reported only as aggregate accuracy, hiding a model that never catches the rare, high-consequence case.
  • A subgroup was silently excluded from effective coverage with no documented scope statement, so users believe the model covers a population it does not.
  • Bias findings exist on paper but no subgroup-sliced evaluation ever appears in the validation record, so the finding was never actually verified as resolved.

How to adapt this risk assessment

  1. Set your document number, owner, and the model or dataset it assesses in the header.
  2. Adjust the likelihood and severity scales in section 2 to match your organization’s existing risk-scoring convention if one exists, rather than running two incompatible scales in parallel.
  3. Add or remove bias categories in section 3 based on the model’s actual data sources; a text-classification model over free-text QMS narratives will weight label bias more heavily than a sensor-driven model will.
  4. Feed the subgroup list from section 5 directly into the model’s validation protocol so sliced evaluation is not left as an afterthought.
  5. Confirm every regulation in section 8 against the current published version before issue.
Use madhadi.com as an app Full screen, works offline, one tap from your home screen.