RxNorm drug data integrity: what medication reconciliation errors mean for AI
RxNorm is the backbone of standardized drug data in the U.S., but inconsistent concept mapping, deprecated codes, and source system fragmentation introduce errors that silently corrupt medication reconciliation. When AI models train on these records without measuring data integrity first, the result is not just bad predictions. It is patient harm at scale.
Medication reconciliation fails before a clinician ever reviews a drug list. It fails at the data layer, where RxNorm concept unique identifiers (RxCUIs) are mapped inconsistently, deprecated without downstream notification, or never assigned at all. A 2023 study in the Journal of the American Medical Informatics Association found that 23% of medication records in a multi-site EHR dataset contained at least one RxNorm mapping discrepancy. That is not a rounding error. That is nearly one in four records carrying a drug identity problem into every system that consumes them.
For AI systems trained on this data, the consequences compound. A clinical decision support tool that ingests unverified medication records does not know whether "metoprolol tartrate 50mg" and "metoprolol succinate 50mg ER" are the same drug or different drugs. RxNorm distinguishes them. But if the source system mapped both to the same RxCUI, or if a deprecated RxCUI was never updated, the AI model inherits a false equivalence. It learns the wrong thing, and it learns it confidently.
This post examines where RxNorm data integrity breaks down, what medication reconciliation errors actually look like in structured data, and why drug data trust is a prerequisite for any AI model that touches medication records.
What RxNorm is and why it matters for drug data trust
RxNorm is the National Library of Medicine's standardized nomenclature for clinical drugs. It provides normalized names and unique identifiers (RxCUIs) for every combination of ingredient, strength, and dose form. If a hospital uses Cerner and a pharmacy uses Epic, RxNorm is what allows a medication list to translate across those systems without losing clinical meaning.
The problem is that translation is only as reliable as the mapping at each endpoint. RxNorm itself is updated monthly, with RxCUIs added, merged, and retired. When a source system fails to update its RxNorm mappings, or when a vendor maps a proprietary drug string to the wrong RxCUI, every downstream system inherits that error.
Three specific failure modes dominate:
What does it mean when it says medication reconciliation?
Medication reconciliation is the process of comparing a patient's current medication list against new orders at every transition of care: admission, transfer, and discharge. The Joint Commission has required it since 2006. The goal is to identify and resolve discrepancies before they cause harm.
But the term obscures the actual complexity. Medication reconciliation is not one conversation between a pharmacist and a patient. It is a data integration problem. The patient's medication list must be assembled from multiple sources: the outpatient EHR, the inpatient pharmacy system, the patient's own recollection, prior discharge summaries, and sometimes external pharmacy benefit manager records. Each of these sources may encode drug data differently.
When the underlying drug identifiers do not agree, reconciliation becomes guesswork. A pharmacist may catch a discrepancy manually. An AI system will not, unless the data has been validated before it reaches the model.
What are the 5 steps of medication reconciliation?
The Institute for Healthcare Improvement outlines five steps:
Every one of these steps depends on drug identity. If the "current list" carries an RxCUI that maps to a retired concept, step three will either miss a discrepancy or flag a false positive. If the "medications to be prescribed" list uses a branded code while the current list uses a generic code for the same drug, the comparison logic breaks unless the system recognizes them as therapeutically equivalent.
What are the 4 types of medication errors?
The National Coordinating Council for Medication Error Reporting and Prevention categorizes medication errors into four primary types:
RxNorm data integrity problems can trigger all four. A prescribing error occurs when clinical decision support fails to flag a duplicate therapy because two RxCUIs that should be linked are not. A dispensing error occurs when a pharmacy system maps a generic drug to the wrong branded equivalent. Administration and monitoring errors follow when the reconciled list itself is wrong.
A 2020 analysis published in BMC Health Services Research found that 46% of medication discrepancies identified during hospital admission involved drug name or dose mismatches traceable to coding inconsistencies rather than clinical intent. The clinician meant the right thing. The data said the wrong thing.
What is an example of medication reconciliation?
Consider a 72-year-old patient admitted to the hospital after a fall. Her outpatient EHR lists "lisinopril 10mg daily" with RxCUI 314076. The hospital's inpatient pharmacy system maps the same drug to RxCUI 197884, which represents lisinopril 10mg oral tablet as a semantic clinical drug rather than a branded formulation.
Both codes refer to the same medication. But an automated reconciliation tool comparing RxCUIs directly will flag this as a discrepancy, requiring manual review. Multiply this by 12 medications on a typical elderly patient's list, and the pharmacist's workload explodes with false positives.
Now reverse the scenario. The patient takes metoprolol succinate ER 50mg, but her primary care EHR mapped it to the RxCUI for metoprolol tartrate 50mg (a different formulation with different pharmacokinetics). The reconciliation tool sees matching RxCUIs and raises no flag. The hospital continues the tartrate formulation. The patient now receives a twice-daily immediate-release drug instead of a once-daily extended-release drug. Her heart rate control deteriorates.
This is not a hypothetical. It is the kind of error that happens thousands of times per year across U.S. hospitals, and it starts with a bad RxNorm mapping.
Key statistics
Why medication reconciliation AI fails without data trust scoring
Several companies are building AI tools for medication reconciliation. They use natural language processing to parse unstructured medication lists, match drug names to RxNorm concepts, and flag potential discrepancies. The intent is good. The architecture is incomplete.
These tools assume the training data is correct. They train on EHR medication records that have already been mapped to RxNorm, treating those mappings as ground truth. But as the statistics above show, the ground truth itself is unreliable in nearly a quarter of records.
An AI model trained on data where metoprolol tartrate and metoprolol succinate are conflated will learn to conflate them. A model trained on records with deprecated RxCUIs will normalize those codes into its predictions. The model does not know its training data is wrong. It cannot know, because no one scored the data before feeding it in.
This is the core problem that SuperTruth's Data Trust Index addresses. The DTI scores every health data record from 0 to 100 across eight dimensions: Provenance (25%), Consent (20%), Recency (15%), Quality (10%), Concordance (10%), Validation (10%), Breadth (5%), and Stability (5%). For medication records specifically, three dimensions carry the most weight.
Provenance determines whether the RxNorm mapping traces back to a verified source system, with a documented chain of custody from the original prescription event to the current coded record.
Concordance measures whether the RxCUI in the record agrees with the same drug's representation in other systems that touch the same patient. If three systems encode the same prescription with three different RxCUIs, concordance drops.
Recency flags whether the RxCUI is still active in the current NLM release or was retired in a prior cycle. A record carrying a deprecated RxCUI scores lower on recency regardless of when the prescription was written.
The cost of ignoring drug data integrity
The financial and clinical costs are stacked. AHRQ estimates that preventable adverse drug events cost the U.S. healthcare system $42 billion per year. A significant portion of those events trace to reconciliation failures during care transitions.
For health systems deploying AI, the liability extends further. An AI tool that recommends a medication change based on corrupted RxNorm data creates a malpractice exposure that did not exist before the tool was deployed. The AI did not cause the data error, but it amplified it into a clinical recommendation.
For pharma companies using RWE for drug safety surveillance, the problem is regulatory. If post-market safety signals are calculated from medication records where 23% contain mapping errors, the signal-to-noise ratio degrades below usefulness. An FDA reviewer asking for the provenance of the underlying drug data will find no answer if no one scored it.
For payers running medication therapy management programs, bad RxNorm mappings mean bad utilization analytics. A drug that appears to be prescribed twice (because it has two different RxCUIs in two different claims) triggers unnecessary prior authorization reviews, costing administrative dollars and delaying patient care.
How to fix RxNorm data integrity before AI training
Four interventions change the trajectory:
1. Score every medication record before model ingestion. The DTI Engine can assess whether an RxCUI is current, properly mapped, and concordant with other records for the same patient. Records scoring below a defined threshold are quarantined for human review rather than passed to a model.
2. Maintain a living RxNorm mapping table. Monthly NLM releases should trigger automated checks against every active RxCUI in the health system's medication database. This is not a one-time cleanup. It is a continuous process.
3. Distinguish term types in reconciliation logic. An SCD (semantic clinical drug) and an SBD (semantic branded drug) for the same molecule are clinically equivalent but informatically distinct. AI reconciliation tools must account for RxNorm's hierarchical relationships, not just string matches or flat code comparisons.
4. Trace provenance to the prescribing event. Every medication record should carry metadata about where the RxNorm mapping was assigned, when, and by what system. Without provenance, there is no way to determine whether a mapping error was introduced at the point of care, during an interface engine transformation, or in a data warehouse ETL process.
The connection between RxNorm, SNOMED CT, and broader clinical data trust
RxNorm does not exist in isolation. Drug data intersects with diagnosis codes (ICD-10), procedure codes (CPT), lab results (LOINC), and clinical findings (SNOMED CT). A medication record that is correct in RxNorm but linked to the wrong ICD-10 diagnosis code creates a different kind of reconciliation error: one where the drug is correctly identified but the reason for prescribing it is not.
This is why data trust must be measured across the entire record, not just within a single terminology. The DTI's concordance dimension specifically evaluates cross-system agreement. If a patient's RxNorm-coded statin prescription does not concordantly link to a SNOMED CT finding for hyperlipidemia or an ICD-10 code for dyslipidemia, the record's trust score drops.
SuperTruth has already mapped this problem across adjacent terminologies. The same data integrity challenges that affect RxNorm mappings affect LOINC code standardization, SNOMED CT mapping, and ICD-10 coding accuracy. Drug data trust is one node in a network of interdependent clinical terminology problems.
Who is responsible for medication reconciliation, and who is responsible for the data underneath it?
Clinically, medication reconciliation is the responsibility of the care team: physicians, pharmacists, and nurses at each transition of care. The Joint Commission holds the organization accountable.
But no one is formally accountable for the integrity of the data that feeds reconciliation. The EHR vendor is not liable for a bad RxNorm mapping that was introduced during a system upgrade. The NLM is not liable for a deprecated RxCUI that a health system failed to update. The AI vendor is not liable for a model trained on unscored medication records.
This accountability gap is where most harm originates. It is also where a trust scoring framework creates the most value. When every medication record carries a DTI score, the question of data quality stops being abstract. A record scoring 38 out of 100 on provenance and concordance is visibly untrustworthy. A record scoring 91 is auditable. The score creates accountability by making integrity measurable.
What comes next for medication AI
The FDA's draft guidance on AI-enabled medical devices will increasingly require documentation of training data quality. CMS is tightening requirements for medication management in value-based care programs. Both trends point toward the same conclusion: unscored drug data will become a regulatory liability.
Health systems that deploy medication reconciliation AI without first scoring the underlying RxNorm data are building on a foundation they cannot audit. When the FDA asks how the training data was validated, "we used whatever was in the EHR" will not be an acceptable answer.
The alternative is straightforward. Score the data first. Enforce a trust floor. Let only records with verified RxNorm mappings, current RxCUIs, and documented provenance reach the model. This is not perfectionism. It is the minimum standard for AI that touches prescribing decisions.
SuperTruth's DTI Engine scores every health data record from 0 to 100 across 8 trust dimensions before your AI model sees it. For medication records, this means flagging deprecated RxCUIs, detecting concordance failures across systems, and verifying provenance back to the prescribing event. If your team is evaluating drug data for AI training, clinical decision support, or regulatory submission, schedule a conversation with the SuperTruth commercial team or (215) 918-4140.
Further reading:
Jason Alan Snyder
Co-founder of SuperTruth and Artists & Robots, and an inventor on the Data Trust Index patents. Twenty-plus years building technology inside Interpublic Group. He writes here nearly every day on data trust, provenance, and what AI should be allowed to act on, and publishes essays on his Substack.
About SuperTruth · LinkedIn · Substack · jasonalansnyder.com
See it in practice
DTI scores the record, not the patient.
8 dimensions. 0 to 100. Travels with every record permanently.