One HCP, six records: why identity resolution is pharma's most expensive data problem

Identity resolution in pharma data systems refers to the process of determining that two or more records across separate data sources refer to the same real-world entity: a specific healthcare professional, a patient, or a product. The concept sounds straightforward. The execution is not, and the consequences of failure run from wasted commercial spend to regulatory breach.

Which identity domains must a pharma CDO resolve?

The pharmaceutical industry maintains three distinct identity domains simultaneously, and each one carries its own stakes.

For healthcare professionals, the commercial stakes are direct. A medical science liaison visiting "Dr. Sarah Chen" at Brigham and Women's Hospital should draw on the same profile as the CRMCRMCustomer Relationship Management: software and strategy to manage and analyse customer interactions throughout their lifecycle.View full definition → record from Veeva, the prescriber data purchased from IQVIA, the speaker bureau record from a third-party compliance vendor, and the adverse event report submitted to FDA. If those records don't resolve to one entity, the company might over-promote to the same physician across three channels (an anti-kickback exposure), under-sample a high-decile prescriber because her records are split across two territories, or, in a pharmacovigilance context, miss a safety signal because duplicate reports are treated as separate cases.

For patients, the stakes shift to safety and privacy together. Longitudinal patient data, drawn from specialty pharmacy dispensing records, HEOR datasets like IQVIA's Real World Solutions or Optum's claims data, and patient support program touchpoints, only yields valid outcomes evidence if the same patient's records are linked across time. Fragmented identity means fragmented therapy timelines, which corrupts both real-world evidence studies submitted to regulators and the internal analytics that guide patient support program design. Misidentification in a pharmacovigilance report can mean a duplicate adverse event inflates a drug's safety signal. The inverse, merging two patients who are distinct, can suppress one.

For products, the problem is less intuitive but equally costly. A single biologic may carry an NDC code in the US dispensing channel, a different identifier in the 340B covered entity system, a CIN from a GPO catalog, and an entirely separate product code in the EU under EMA's IDMP (Identification of Medicinal Products) framework. Failure to resolve these to a single product identity means government price reporting, specifically ASP and AMP calculations for Medicaid and Medicare, draws from mismatched product records. The financial exposure from that failure is not theoretical: CMS audits have resulted in significant manufacturer repayments, and false claims liability follows.

How do pharma teams match HCP records across sources?

The mechanics fall into three broad approaches: deterministic matching, probabilistic matching, and graph-based resolution.

Deterministic matching is exact: two records that share an NPI number (for US HCPs) or a DEA number refer to the same entity. Full stop. NPPES, the National Plan and Provider Enumeration System, is the canonical US source. Pharma data teams that build their HCP master on NPI as the primary key and enrich from there, via OneKey from IQVIA or DDD from Wolters Kluwer Health, are working with the cleanest starting point available. The limit is obvious: NPI only covers US prescribers, does not exist for key opinion leaders who don't prescribe (lab researchers, KOL-type PhDs), and provides no cross-border linkage for global companies tracking HCPs across the EU.

Probabilistic matching activates when deterministic keys are absent or conflicting. The algorithm scores record pairs on weighted attributes: surname, given name, specialty, address, phone, graduation year, affiliated institution. A Jaro-Winkler string distance on "Catherine Moreau" versus "C. Moreau-Dubois" at the same hospital address in Lyon might return a match score of 0.87. The matching engine applies a threshold, say 0.80, to declare a match, a lower threshold to flag for human review, and below that to treat records as distinct.

A concrete example: a European pharma company launches a new oncology biologic across France, Germany, and the Netherlands. Their medical affairs team has 14,000 HCP records from three national distributors, a conference attendance list from ESMO 2025, and a speaker engagement database. The same oncologist, Dr. Moreau, appears four times across those sources, spelled differently in each, affiliated with two hospitals (she holds a joint appointment), and with a French RPPS identifier in one record and an email address in another. A deterministic pass matches two of the four records via RPPS. The probabilistic layer resolves the third via name-address similarity. The fourth, from the ESMO list, carries only her name and an institutional email from her secondary affiliation: it goes to a human review queue. That workflow, deterministic first, probabilistic second, human adjudication third, is the production standard at companies like Roche and Sanofi for HCP master data maintenance.

Graph-based resolution, increasingly adopted as of 2026, models entities and their relationships as nodes and edges. Rather than comparing record pairs in isolation, it asks: does this unresolved HCP record share an institutional affiliation, a co-author relationship, and a prescribing ZIP code cluster with a known entity in the graph? Companies running identity resolution on graph databases (Neo4j is common in this space) find thatthe golden record emerges from relationship evidence, not just attribute similarity, which substantially improves recall for high-mobility HCPs like academic physicians who rotate across institutions.

Accuracy thresholds for pharmacovigilance, commercial targeting and RWE

Identity resolution is not uniformly worth solving to the highest fidelity across every use case.

For pharmacovigilance and ICSR (Individual Case Safety Report) management, there is no acceptable error rate. FDA regulations under 21 CFR Part 314 and EMA's GVP Module VI require accurate patient and reporter identification. A duplicate that doubles an adverse event count for a compound can trigger a regulatory safety review. Invest in deterministic resolution, mandatory human review queues, and a well-governed survivorship policy here without compromise.

For commercial targeting, probabilistic matching to 85-90% accuracy is usually good enough. A 5% error rate in a Veeva CRM HCP universe means some reps visit the wrong decile tier or miss a newly important prescriber. That matters commercially, but it does not create legal exposure by itself, unlessthe errors feed into speaker bureau selection or grant decisions, where anti-kickback scrutiny applies.

For real-world evidence studies, the threshold depends on the regulatory pathway. An RWE study supporting a label extension submitted to FDA under 21st Century Cures Act provisions will face methodological scrutiny on patient matching quality. An internal HEOR analysis informing payer contracting strategy has more latitude.

The honest tradeoff is this: enterprise-grade identity resolution requires sustained investment in data stewardshipdata stewardshipA business-side owner responsible for the quality, consistency and appropriate use of data in their domain.View full definition →, not just tooling. Vendors like IQVIA, Veeva, and Komodo Health sell pre-resolved HCP and patient universes as products, which reduces build cost but transfers governance risk. When their matching is wrong, correcting it inside your own systems without overwriting the vendor's primary key is a non-trivial architectural problem.

The CDOs who get this right treat identity resolution as infrastructure owned by data governancedata governanceData governance is the set of policies, roles, and processes that ensure data is accurate, secure, well-defined, and used responsibly across an organization.View full definition →, not a feature owned by a single commercial application. That distinction determines whether the golden record holds across a product launch, a merger, and a regulatory inspection.

The full course on this sector:Data in Pharmaceuticals.