Medical AI trained on claims data needs three checks before use

The International Medical Device Regulators Forum (IMDRF) finalized 10 good machine learning practice principles for artificial intelligence (AI) in medical devices on January 27, 2025. The principles treat representative data, independent testing, human-AI evaluation and post-deployment monitoring as parts of one product lifecycle.

Health insurance claims and other administrative health records can support machine-learning research when the governing rules permit secondary use. Access alone does not answer whether the fields and labels fit the target task or whether the resulting model works in its intended clinical setting.

First check: Permission covers a defined use

The Organisation for Economic Co-operation and Development (OECD) distinguishes de-identified records from anonymous data and calls for a defined legal basis, review and approval procedures, and safeguards for secondary uses of personal health data. Its recommendation is a policy standard for governments, not a common legal authorization that applies in every country.

The recommendation says consent should be informed and withdrawable when consent is the basis for processing, but it also recognizes lawful alternatives and exemptions. The applicable country's rules and the dataset's access terms determine which path applies. A project record should identify the authorized purpose, fields, recipients, processing environment, retention period, permitted outputs and consequences of a later withdrawal or restriction.

Second check: Claims data fit the target task

A 2023 systematic review examined 70 machine-learning studies using administrative health records, describing the records as well-structured categorical data that can show how patients interact with a health system. The authors also identified persistent limits in the data and the methods used to analyze them.

The review explains that administrative codes record transactions and may describe a person's health state only indirectly. It identifies missing or incorrect values, inconsistent formats and fragmented records as recurring problems, while a single data source may leave parts of a person's care unobserved.

A separate 2023 systematic review of 103 papers identified completeness, correctness, concordance, plausibility, currency, conformance and bias as recurring dimensions of electronic health record (EHR) quality, while finding no standard assessment method. That review concerned EHRs rather than insurance claims, so its dimensions are prompts for assessment rather than evidence that a claims source is fit.

A training review needs a data dictionary, coverage dates, coding changes, missing-value patterns, label definitions and a record of which population each source represents. When a billing code serves as a proxy for a clinical outcome, the team needs to test that proxy against a reference suited to the intended use. A large row count cannot repair a field that measures the wrong event.

Third check: Clinical evidence matches the intended setting

IMDRF says clinical evaluation datasets should represent the intended patient population, use environment and measurement inputs. Its principles also call for training and test sets to remain appropriately independent, with external validation proportionate to risk.

The World Health Organization's 2021 guidance calls for testing under intended conditions, checks for performance differences across relevant groups and independent oversight of evaluation. It also calls for performance and harm to be monitored during use, with a process for stopping systems that are ineffective.

Federated learning can keep records at participating hospitals while exchanging model updates, but it does not establish data quality or clinical performance. Whether records move or remain on site, each intended deployment still needs evidence covering its patients, inputs, users, workflow and failure response.

Deployment adds monitoring and change control

Joint 2023 principles from the US Food and Drug Administration (FDA), Health Canada and the United Kingdom's Medicines and Healthcare products Regulatory Agency say a change plan should define planned modifications, verification and validation methods, impact assessment, and mechanisms to stop or revert a change that misses its performance criteria. The document says these plans may be implemented differently across jurisdictions, so it is not a global authorization route.

A deployment file should connect the deployed model version to its data lineage, validation results, monitoring measures, investigation owner and rollback process. If the input mix, coding rules or clinical workflow changes, the team needs to determine whether the earlier evidence still describes the deployed system. Availability, regulatory status and reporting duties vary by country.

A three-check record keeps the decisions separate

A compact review can keep three linked records. The governance record identifies authority, purpose, recipients and controls. The data record documents provenance, definitions, missingness, representation and labels. The clinical record connects an independent test, intended workflow, human authority, deployed version and monitoring response.

Each record can fail independently. A valid data-use decision cannot correct a wrong label, a clean dataset cannot replace external testing, and an accuracy score cannot allocate responsibility or set update rules. Approval to enter a controlled data environment is therefore the start of an evidence process, not the end.

The cited sources define governance principles and recurring data-quality problems; they do not establish that any national insurance database is available to a commercial developer or that a model trained on it is authorized for clinical use in any country. APPI News could not find a published cross-country dataset showing how many insurers permit their claims data to train commercial medical AI at the time of writing.

Frequently asked questions

Can de-identified claims always be used to train AI?
No. The OECD distinguishes de-identified data from anonymous data, and the applicable country, dataset owner and authorized purpose determine the legal basis and safeguards.

Does a national claims dataset represent every patient?
No. Coverage depends on who is insured, which services generate a claim, how codes are recorded and what care occurs outside the dataset. Representation has to be tested against the population and setting named for the model.

Does a strong score on claims data make a model ready for a hospital?
No. The model still needs independent evidence in clinically relevant conditions, a tested human workflow and monitoring tied to the deployed version.