Topic 4.9 — Data Quality & Master Data Management

In health and care, a wrong NHS number, a missing allergy or a duplicate patient record is not a data-tidiness problem but a clinical hazard, a payment error and a research flaw, so data quality and master data management must be run as managed disciplines with named owners — not left to chance at the point of capture.

Why this matters in health and care#

Every clinical decision rests on data someone recorded earlier: an allergy flagged, a result filed, a medication reconciled, an identity confirmed. When that data is accurate, complete and attached to the right person, care flows safely. When it is wrong, absent or attached to the wrong person, the system does exactly what it is told — and harms someone. A duplicate record hides a critical allergy on the chart the clinician is not looking at; a mis-keyed NHS number sends a result to the wrong patient's file; a stale address means a cancer screening invitation never arrives. These are not hypothetical: patient-safety incidents involving mis-identification and record confusion are a recurring theme in national safety reporting.

Data quality is therefore a clinical-safety issue before it is an analytical one (see Topic 1.6 — Clinical Safety & Risk Management). It is also a question of public money and fairness. The NHS pays providers, plans services and holds them to account on the basis of what the data says happened; if activity is under-recorded or mis-coded, the wrong provider is paid the wrong amount and commissioning decisions rest on a distorted picture. Population health analytics, waiting-list validation and health-inequalities work all inherit the quality of the underlying records (see Topic 4.5 — Data, Analytics & Population Health Management).

The stakes rise sharply as data is reused. A record captured once at a busy front desk may later train an algorithm, justify a business case, feed a national dataset and settle a payment dispute. The old maxim — garbage in, garbage out — has become garbage in, harm out, because poor data now propagates automatically and at scale into artificial-intelligence models and automated decisions (see Topic 8.4 — AI & Emerging Technology Governance). And master data — the shared, slowly-changing facts about who patients, staff, organizations and medicines are — is the connective tissue that lets dozens of systems agree they are talking about the same person. Get the master data wrong and every downstream system inherits the confusion.

Core concepts#

Data quality is the degree to which data is fit for its intended uses in operations, decision-making and planning. It is conventionally assessed across a set of dimensions. Accuracy is whether the data correctly describes the real-world thing it represents; completeness is whether the required data is present; timeliness is whether it is available when needed and reflects the current state; validity is whether it conforms to the defined format, type and range; consistency is whether it agrees with itself and with other sources; and uniqueness is whether each real-world entity appears once and only once. A record can be valid (correctly formatted) yet inaccurate (wrong), so the dimensions are complementary, not interchangeable.

Master data management (MDM) is the discipline by which business and technology together ensure the uniformity, accuracy, stewardship and accountability of an organization's shared master data — the core entities that many systems reference: patients, staff, organizations, sites, and medicines. The aspiration of MDM is a golden record: a single, trusted, reconciled version of each entity, assembled from many source systems, that acts as the authoritative reference. This is closely related to the idea of a single source of truth — the practice of mastering each data element in exactly one authoritative place so that everything else refers back to it rather than keeping its own divergent copy.

Reference data is a specific, important sub-type. Reference data is data used to classify or categorize other data — code lists, value sets, organization codes, units of measure — that is typically defined externally and changes slowly. In health and care this includes clinical terminologies such as SNOMED CT and national organization and site codes; the standards for these are owned nationally and mandated as information standards (interoperability and exchange are covered in Topic 4.3 — Interoperability & Data Standards, which this topic does not repeat).

The NHS number is the primary unique patient identifier in England, Wales and the Isle of Man — a ten-digit number with a checksum that lets systems detect a mistyped digit. It is the anchor for patient master data. Attaching the right NHS number to the right person is a record linkage problem — finding records across different sources that refer to the same entity — solved in England through the Personal Demographics Service (PDS), the national demographic index, and locally through a Master Patient Index (MPI). When linkage fails you get duplicates (one person, many records) or overlays (two people merged into one), both dangerous.

The operating disciplines that keep quality high are data governance — the framework of decision rights and accountabilities for data — and the data steward, the named role responsible for the quality and fitness of a defined data domain. Two techniques recur: data profiling, examining data to gather statistics about its actual condition, and data cleansing, detecting and correcting corrupt, inaccurate or duplicate records. Prevention beats cure: data validation at the point of capture — validation rules and constraints that reject or query bad input before it enters the system — is cheaper and safer than cleansing it later.

Best practices#

1. Define data quality in the specific dimensions that matter to each dataset, and measure them. "Improve data quality" is not an objective; "raise NHS number completeness on the emergency dataset above the national threshold" is. Agree which of accuracy, completeness, timeliness, validity, consistency and uniqueness matter most for each critical dataset, set measurable targets, and report against them routinely. The NHS Data Quality Maturity Index (DQMI) gives a ready-made, comparable measure for many national datasets; use it as a baseline and track your own local metrics for the fields clinicians and finance actually depend on.

2. Validate at the point of capture, not in a downstream clean-up. The cheapest error to fix is the one that never enters the system, and the person best placed to correct a record is the one creating it, while the patient is in front of them. Build validation rules, mandatory fields, controlled pick-lists and NHS-number checksum verification into the forms and interfaces staff actually use, and confirm identity against the Personal Demographics Service at registration. Every error caught at capture is an error that never has to be found, investigated and reversed downstream — after it may already have caused harm.

3. Treat the NHS number as a safety-critical anchor and verify it everywhere. The NHS number is the single most valuable identifier in the system because it is the key that links a person's records safely across organizations. Trace it against the Personal Demographics Service, validate its checksum on entry, and flag records that lack a verified number rather than letting unmatched data flow on. A result, referral or prescription attached to the wrong number is a clinical hazard, so treat NHS-number quality as a patient-safety control, monitored and escalated like any other.

4. Establish master data management with a golden record for each core entity. Decide deliberately which system masters each entity — patients, staff, organizations, medicines — and how a single reconciled golden record is assembled and maintained from the source systems. Run a Master Patient Index reconciled against the national demographic service, and manage a controlled, auditable process for merging duplicates and un-merging erroneous overlays, because both operations touch live clinical records and can themselves cause harm if done carelessly. Extend the same rigour to staff and organization master data, which underpins access control, payment and reporting (see Topic 4.2 — Identity & Access Management, and Topic 5.4 — Enterprise Resource Planning (ERP), which owns corporate/finance master data).

5. Name data owners and data stewards, and give them real authority. Data quality decays without clear accountability, so assign each critical data domain a senior data owner accountable for it and an operational data steward who monitors, investigates and drives correction. Stewards need standing, tooling and time — not a line added to an already-full job — and a route to escalate systemic problems (a broken form, a confusing workflow, an interface that drops fields) to the people who can fix the cause. Ownership converts data quality from everybody's vague concern into somebody's explicit responsibility.

6. Profile before you cleanse, and cleanse at the source. Before any improvement or migration, profile the data to understand its true state — how many records lack an NHS number, how many duplicates exist, which fields are systematically miscoded — so you target effort where it counts rather than guessing. When you cleanse, fix the record in the system that masters it and address the process that produced the error, not just the copy in a downstream warehouse, or the same defect will simply reappear. One-off clean-ups that ignore the cause are a treadmill; the goal is to stop the tap, not just mop the floor.

7. Adopt national information standards and reference data rather than inventing local codes. Consistency across organizations is only possible when everyone classifies the world the same way, using nationally-defined standards mandated through the Data Alliance Partnership Board (DAPB) process (formerly SCCI/ISB standards) and terminologies such as SNOMED CT. Manage reference data and code lists under version control with a clear update process, because a silently changed value set breaks trend analysis and can corrupt downstream logic. Local extensions should be the rare, governed exception, mapped back to the national standard (see Topic 4.3 — Interoperability & Data Standards).

8. Make data quality visible to the people whose decisions depend on it. Data quality improves when it is seen — publish quality metrics to clinical, operational and finance teams so that the cost of poor data lands where the work is done, not buried in a data team. Show downstream consumers the provenance and known limitations of the data they are using, so an analyst, a payment team or a model-builder can judge whether it is fit for their purpose (see Topic 10.1 — Key Performance Indicators (KPIs)). Transparency about quality builds the trust that makes people use data at all, and the discomfort of a visible red metric is often what finally funds the fix.

Questions to discuss with your team#

1. Which specific data errors in our organization could most plausibly harm a patient, and how would we even know they were happening? This reframes data quality from an abstract virtue into a safety question with named hazards, which is the only version that gets prioritized. Work through the concrete failure paths: a duplicate record that hides an allergy, a mis-verified NHS number that mis-routes a result, a stale contact detail that means a two-week-wait referral never reaches the patient. The hard part is detection — most data errors are silent, discovered only when harm occurs or an audit stumbles on them, so debate whether you have any proactive way of finding them or are relying on luck and complaints. A good, honest answer names your two or three highest-harm data hazards, states plainly whether you could currently detect each one, and treats the gaps as clinical-safety risks to be logged and owned (see Topic 1.6 — Clinical Safety & Risk Management) rather than as tidiness to be got to eventually.

2. Where is the single source of truth for our core entities — patients, staff, organizations, medicines — and do our systems actually respect it? Many organizations discover, when they look, that the same patient exists differently in the patient administration system, the emergency system, the pathology system and three departmental databases, with no agreement on which is authoritative. Discuss, entity by entity, which system is supposed to master it, how a golden record is assembled, and how duplicates and overlays are found and resolved — including who is allowed to merge two records, since that operation can itself cause harm. Surface the tension between local autonomy (departments that want their own copy they can control) and system-wide safety (one reconciled truth), and be honest that master data management is continuous stewardship, not a one-off migration. A strong answer identifies the mastering system for each entity, admits where today there is no agreed source of truth, and commits to a named owner and a reconciliation process rather than a database clean-up that will silently rot again.

3. When our data feeds a payment, a national return, a business case or an AI model, whose job is it to know whether it is fit for that purpose? Data captured for one reason is constantly reused for others, and quality that is adequate for a clinical note may be dangerously misleading in an algorithm or a funding formula. Debate who is accountable for judging fitness at the point of reuse: does the analyst, the finance lead or the model-builder know the provenance, completeness and known defects of the data they are relying on, or do they assume it is clean because it looks tidy? Consider concrete cases — a coding gap that quietly understates activity and therefore income, or a demographic field with biased missingness that skews an AI model against a patient group (see Topic 8.4 — AI & Emerging Technology Governance, and Topic 1.8 — Data Ethics, Consent & Public Trust). A good answer establishes that fitness-for-purpose is assessed explicitly at each point of reuse, that data carries visible provenance and quality caveats, and that no one is silently trusting numbers whose limitations they have never seen.

4. Are we preventing errors at the point of capture, or quietly pushing the burden of data quality onto already-stretched clinicians? Prevention at source is the cheapest and safest form of data quality, but it lives or dies in the design of the forms and workflows staff actually use under pressure. Discuss whether your capture screens genuinely help — checksum validation, controlled pick-lists, PDS-verified identity, sensible defaults — or whether they instead pile mandatory fields and warnings onto a busy registration desk until people learn to click past them. There is a real tension here: too little validation lets bad data flow on, but clumsy validation breeds workarounds, dummy values and alert fatigue that corrupt the data in subtler ways. Look at where staff are entering "not known", reusing a default, or keeping a private spreadsheet to avoid a painful workflow, because each is a signal that the system is fighting the work rather than supporting it. An honest answer treats capture design as a joint clinical, operational and data problem, tests it with the people who use it, and measures whether validation actually reduces downstream errors rather than just relocating them.

5. Do the people we call data owners and stewards have the standing, time and tools to change anything, or is stewardship a title bolted onto an already-full job? Data quality decays without clear accountability, but naming an owner on a slide is not the same as resourcing one, and stewardship fails quietly when the named person has neither authority nor capacity. Debate what your stewards can actually do when they find a systemic problem — a broken interface that drops fields, a confusing form, a workflow that manufactures duplicates: can they get it fixed, or can they only log it and hope? Surface the trade-off between spreading stewardship thinly across many people who each do a little in the margins, and concentrating it in a smaller number of genuinely resourced roles with time carved out and tooling provided. Consider whether stewards have a real escalation route to the people who own the causing systems and processes, and whether senior data owners are held to account for the quality of their domain. A good answer is honest about whether stewardship is currently real or ceremonial, and commits to authority, time and tooling rather than to another unfunded responsibility.

6. How do we manage national standards and reference data over time, and what happens when a code list or terminology changes underneath us? Reference data and terminologies such as SNOMED CT are defined externally and change slowly but continually, and a silently updated value set can break trend analysis or corrupt downstream logic without anyone noticing. Discuss whether your code lists and mappings are under version control with a clear update process, or whether they live in undocumented lookups that someone edits in place. Examine the pull towards inventing local codes when the national standard feels awkward, and be honest about the long-term cost: local classifications make your data impossible to compare or aggregate with anyone else's and become a permanent mapping burden (see Topic 4.3 — Interoperability & Data Standards). Consider how you would even detect that a reference value changed, and how you would assess the impact on the analyses, payments and models that depend on it. A strong answer treats reference data as a governed, versioned asset with a named process for adopting national changes, keeps local extensions rare and mapped back to the standard, and can trace which downstream uses a change would affect.

In practice: a health & care example#

Northmoor Integrated Care System runs a shared care record spanning two acute trusts, a mental-health trust, community services and 60 GP practices. After a serious incident — a patient's severe penicillin allergy recorded in one trust was invisible in another because the record had failed to link, and the patient was prescribed the drug — the board treats data quality as a clinical-safety priority rather than a back-office concern. A rapid review finds the root cause is master data: the same patients exist as duplicates across systems, and NHS-number verification is inconsistent at the front door.

The ICS establishes a data governance forum with a named senior data owner for patient master data and operational data stewards in each provider. Their first act is to profile the estate: they discover that NHS-number verification against the Personal Demographics Service happens reliably in some registration workflows and not others, and that the shared record contains a significant population of duplicate and unmatched records. Rather than a mass clean-up, they fix the cause first — enforcing PDS verification and checksum validation at every point of capture — so the duplicate population stops growing while they work through the backlog.

They stand up a Master Patient Index reconciled against the national demographic service, with a controlled, audited, clinically-supervised process for merging duplicates and reversing any erroneous overlays, recognizing that both operations touch live records. They adopt the NHS Data Quality Maturity Index as an external baseline for their national dataset submissions and add local metrics for the fields clinicians most depend on — allergies, NHS-number completeness, key demographics — publishing them to each provider's clinical and operational leaders so the numbers cannot be ignored.

Eighteen months later the change is visible in outcomes, not just dashboards: NHS-number verification is near-universal at capture, the duplicate rate is falling, and allergy information reconciles reliably across the shared record. When the ICS later scopes a population-health risk-stratification model, the analytics team can state the provenance and completeness of the underlying data and flag where missingness might bias the model, because the master data and quality metrics now exist to support that judgement (see Topic 4.5 — Data, Analytics & Population Health Management). The unglamorous work — a governance forum, stewards, validation at capture, an MPI — converted an invisible, harmful risk into a managed one.

Sector lenses#

Startup#

A digital-health startup often treats data quality as a future problem, focusing on features while accepting messy data to move fast. That is a mistake with health data, where a mis-matched patient or an unverified NHS number is a safety issue from the first live patient, not the thousandth. Build validation, NHS-number checksum verification and — where you can access it — Personal Demographics Service matching into your capture flows from the start, because retrofitting identity discipline after you have a large corpus of unmatched records is painful and risky. Your credibility with NHS buyers depends on being able to show that your data is trustworthy, its provenance is clear, and you will not pollute their master data when you integrate.

Small business#

An established small provider — a GP practice, a community pharmacy, a care home or a single-site clinic — carries the same duty to get identity and allergies right as a large trust, but with no data team and often no dedicated IT. Here master data management is not a platform project; it is the disciplined use of the tools you already have: verifying the NHS number against the Personal Demographics Service at registration, keeping demographics current, and not letting duplicate records accumulate in your clinical system. Lean on the national infrastructure that is provided for you — PDS, SNOMED CT, national organization codes — rather than inventing local shortcuts, because it does the heavy lifting a small organization cannot. Name one person as the practical owner of data quality even if it is part of a wider role, and use your system supplier's validation, matching and de-duplication features rather than tolerating messy data because "we're too small to worry about it" — the harm from a wrong record is the same size whatever the organization.

Enterprise#

A large NHS trust or ICS runs dozens of systems that each hold their own version of patients, staff and organizations, so its central challenge is reconciliation and stewardship at scale. The realistic goal is a governed Master Patient Index and a small number of authoritative golden records, enforced through interfaces and validation, not a single monolithic database that replaces everything. Invest in named data owners and stewards with real authority, tie master data to identity and access management and to payment, and treat merges and overlays as clinically-supervised operations because at this scale they touch live care. The biggest wins come from prevention at capture and disciplined national-standard adoption applied consistently, rather than heroic periodic clean-ups that decay the moment they finish.

Government#

A national body — NHS England, and the standards bodies that succeeded NHS Digital — shapes the data-quality environment every provider works in. It runs the shared infrastructure that makes local quality possible: the Personal Demographics Service as the national demographic index, the NHS number as the common key, national terminologies and reference data, and the Data Quality Maturity Index that makes performance comparable across organizations. Its leverage is to mandate information standards through the Data Alliance Partnership Board process, to publish comparable quality metrics that create improvement pressure, and to steward reference data so that everyone classifies the world the same way. Government also carries the equity duty: monitoring for biased missingness and ensuring data quality does not silently disadvantage groups whose records are least complete (see Topic 9.1 — Working with Local & National Governments).

Common failure modes#

• Cleansing forever, fixing nothing. Repeated downstream clean-ups that never address the capture process are a treadmill; the same duplicates and errors reappear each cycle. Fix the cause at the source.
• Valid but wrong. A record that passes format validation can still be inaccurate — a correctly-formatted but wrong NHS number is more dangerous than a blank one because it looks trustworthy. Validate against real reference sources, not just format.
• No named owner. When data quality is everyone's concern it is no one's job; problems are admired in meetings and never resolved. Assign accountable owners and empowered stewards.
• Master data by accident. Letting each system keep its own divergent copy of patients or organizations, with no agreed source of truth, guarantees inconsistency and dangerous duplicates. Decide deliberately what masters each entity.
• Reckless merges and overlays. Merging two records or splitting a wrongly-merged one touches live clinical data; done without control or clinical oversight it can hide or misattribute critical information. Make it an audited, supervised process.
• Local codes that break comparability. Inventing local classifications instead of using national standards makes data impossible to compare or aggregate and corrupts trend analysis when they change silently. Adopt and version national reference data.
• Silent reuse. Feeding data of unknown quality into payments, national returns or AI models without stating its provenance and limitations propagates hidden errors at scale. Make fitness-for-purpose an explicit judgement at each point of reuse.

Maturity model#

Checklist#

• Critical datasets have agreed quality dimensions, measurable targets, and routine reporting, benchmarked where possible against the DQMI.
• NHS number is checksum-validated and verified against the Personal Demographics Service at every point of capture.
• Each core entity (patient, staff, organization, medicine) has a designated mastering system and a golden record.
• A Master Patient Index is reconciled against the national demographic service, with an audited, clinically-supervised merge and overlay process.
• Every critical data domain has a named data owner and an empowered data steward with tooling and an escalation route.
• Data is profiled before migration or improvement, and cleansing fixes the source system and the causing process, not just downstream copies.
• National information standards and terminologies (via the DAPB process, SNOMED CT) are adopted, and reference data is version-controlled.
• Data quality metrics are published to the clinical, operational and finance teams whose decisions depend on the data.
• Provenance and known quality limitations travel with data when it is reused for payment, national returns, research or AI.
• High-harm data-quality hazards are logged and owned as clinical-safety risks, with a means of proactive detection.

Key sources#

• NHS England / NHS Digital — Data Quality Maturity Index (DQMI) — a monthly comparable measure of data quality across national datasets.
• NHS England — Personal Demographics Service (PDS) and NHS number guidance — the national demographic index and primary patient identifier.
• NHS England — Information standards published through the Data Alliance Partnership Board (DAPB) process (formerly SCCI/ISB standards).
• DAMA International — DMBOK (Data Management Body of Knowledge) — the recognised framework for data governance, quality and master data management.
• SNOMED International / NHS England — SNOMED CT clinical terminology and dm+d (dictionary of medicines and devices) for medicines master and reference data.
• HM Government — Data Standards Authority and the Government Data Quality Framework — cross-government guidance on data quality dimensions and stewardship.
• National Data Guardian — Caldicott principles — on the appropriate use and quality of confidential patient information.

References#

1. Data quality — Wikipedia — https://en.wikipedia.org/wiki/Data_quality
2. Master data management — Wikipedia — https://en.wikipedia.org/wiki/Master_data_management
3. Single source of truth — Wikipedia — https://en.wikipedia.org/wiki/Single_source_of_truth
4. Reference data — Wikipedia — https://en.wikipedia.org/wiki/Reference_data
5. SNOMED CT — Wikipedia — https://en.wikipedia.org/wiki/SNOMED_CT
6. NHS number — Wikipedia — https://en.wikipedia.org/wiki/NHS_number
7. Record linkage — Wikipedia — https://en.wikipedia.org/wiki/Record_linkage
8. Data governance — Wikipedia — https://en.wikipedia.org/wiki/Data_governance
9. Data steward — Wikipedia — https://en.wikipedia.org/wiki/Data_steward
10. Data profiling — Wikipedia — https://en.wikipedia.org/wiki/Data_profiling
11. Data cleansing — Wikipedia — https://en.wikipedia.org/wiki/Data_cleansing
12. Data validation — Wikipedia — https://en.wikipedia.org/wiki/Data_validation
13. NHS England — Data Quality Maturity Index (DQMI) — NHS England — https://digital.nhs.uk/data-and-information/data-tools-and-services/data-services/data-quality
14. NHS England — Personal Demographics Service (PDS) — NHS England — https://digital.nhs.uk/services/personal-demographics-service
15. NHS England — Information standards and the Data Alliance Partnership Board — NHS England — https://www.england.nhs.uk/digitaltechnology/digital-primary-care/information-standards/
16. DAMA International — Data Management Body of Knowledge (DMBOK) — DAMA International — https://www.dama.org/
17. HM Government — The Government Data Quality Framework — GOV.UK — https://www.gov.uk/government/publications/the-government-data-quality-framework
18. National Data Guardian — Caldicott Principles — GOV.UK — https://www.gov.uk/government/publications/the-caldicott-principles