Privacy and preparation
Healthcare administrative dataset sample review: what to inspect
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A privacy reviewer should inspect a healthcare administrative data sample for five things: that the task is administrative and nonclinical, that the record holder has a documented permission for this use, that every field is necessary for the stated workflow, that identifiers in coded fields and free text are handled by a named method, and that specially protected records such as substance use disorder program files are excluded or escalated.
Redaction
Healthcare administrative
Key takeaways
- Tie the sample to one administrative task and keep clinical narrative outside it.
- Record who holds the records and which agreement governs this use before reviewing fields.
- Map every field to a workflow step; drop or generalize anything that has no step.
- Expect identifiers in claim numbers, dates, correspondence and call notes, not only in name fields.
- Stop and escalate when records from separately protected programs, such as substance use disorder treatment, appear.
What should privacy reviewers inspect in a healthcare administrative data sample?
They should inspect whether the sample matches its stated administrative task, whether the permission to use it is documented, whether each field earns its place, whether identifiers are handled by a named method, and whether any record belongs to a category that needs separate review. The sample is the cheapest place to find these problems; the same problems found after a full extract is prepared cost far more to fix.
This page is about administrative records: eligibility checks, claim status, remittance and denial codes, prior authorization routing, scheduling and billing correspondence. It does not cover clinical records, clinical decisions or any advice about care. Where the Health Insurance Portability and Accountability Act (HIPAA) applies, the reviewer works from the HHS guidance cited here and from counsel; this page is general information, not a legal opinion.
The protocol below assumes a supplier, such as a revenue cycle company, clearinghouse, billing service or provider group, is reviewing a sample of its own records before deciding whether a larger extract can be prepared for an AI developer.
Where is the nonclinical boundary for an administrative workflow sample?
The boundary sits at the information the administrative decision actually used. A claim status agent needs payer, claim identifiers, dates of service, status and adjustment reason codes; it does not need the clinical narrative that justified the service. A prior authorization router needs the request type, the payer's required documents and the decision status, not the content of the clinical attachments.
Administrative data in the U.S. often follows HIPAA transaction standards. CMS lists the adopted standards, including ASC X12N 837 for claims, 270/271 for eligibility inquiry and response, 835 for payment and remittance advice and 278 for referral and prior authorization. Naming the transaction set behind each sample field helps the reviewer see where clinical content can ride along, for example in diagnosis and procedure code segments or in attachments that a 278 request points to.
| Workflow | Administrative fields usually needed | Clinical detail to keep outside the task |
|---|---|---|
| Eligibility verification | Payer, plan type, coverage dates, benefit category, response status | Diagnoses, reasons for the visit |
| Claim status follow-up | Claim and payer identifiers, service dates, status codes, adjustment reason codes | Clinical notes, test results |
| Denial and appeal routing | Denial reason code, appeal deadline, payer, appeal outcome | Medical necessity narrative unless the task is scoped to it and separately reviewed |
| Prior authorization routing | Request type, payer requirement list, submission and decision dates, status | Content of clinical attachments |
| Patient billing correspondence | Statement type, balance category, contact outcome, dispute flag | Treatment descriptions in letters or call notes |
Whose permission covers this use of the records?
Start with the holder's role, because it decides what the holder may do. HHS describes a business associate as a person or organization outside a covered entity's workforce that performs functions involving protected health information on the covered entity's behalf, and lists claims processing or administration, data analysis, billing and practice management among those functions. A business associate's uses are limited by its agreement with the covered entity.
The reviewer should therefore record four separate answers: who has access to the records, what privacy rules and notices permit, who owns the records, and what contracts allow by way of licensing. Access through a billing system is not permission, and de-identification alone does not create a right to license records that a business associate agreement or client contract reserves to someone else. Counsel reviews the agreements before the reviewer goes further.
Is every field in the sample necessary for the stated task?
Each field should be traceable to a step in the workflow, and anything that is not should be dropped before the full extract. HHS guidance on the minimum necessary standard says covered entities must make reasonable efforts to use, disclose and request only the minimum amount of protected health information needed for the purpose, and that routine, recurring disclosures can follow standard protocols that set those limits in advance.
A field necessity table makes this auditable. List every column and attachment type, the workflow step that uses it, what happens to model quality if it is removed, and the decision: keep, generalize, tokenize or drop. Free-text fields deserve their own row each, because a call note column can carry more sensitive information than the rest of the table combined.
Generalizing is often better than dropping. A service date can become a month or a week offset from the claim's first event, a ZIP code can become a state or region, and a payer name can become a payer type when the task does not depend on the specific payer. Each of these keeps the workflow signal the model needs while removing detail the task never used. Record the transformation in the table so a buyer knows what the field now means.
The owner of the workflow, not the privacy reviewer, should defend each kept field. When the owner cannot name the step that uses a field, the default is to remove it; when the owner names a step but the field is sensitive, the reviewer asks whether a coarser version would serve the same step.
The sample review protocol and evidence checklist
Run the protocol on a random sample drawn across the full date range and every source system, not on records someone picked because they looked clean. Each step records evidence and has a stop condition; a stop is not a failure of the review but its main output.
Keep the sample separate from the eventual extract and do not reuse it as training data without the same review the full extract will get. The sample's job is to reveal the shape of the problems; once the field table and stop rules are settled, the reviewer should repeat steps 5 to 7 on a fresh sample from the prepared extract, because transformations can introduce new problems such as tokens that still sort by patient name.
- Evidence checklist: task statement; holder role and agreement references; privacy notices in force; sampling frame and seed; field necessity table; identifier inventory; free-text finding log; special-category flags; de-identification method memo; payer and transaction version for each source; reviewer sign-off with date and conditions.
| Step | What the reviewer does | Evidence recorded | Stop or escalate when |
|---|---|---|---|
| 1. Task statement | Confirms the sample maps to one administrative task | Signed task statement with version | The task needs clinical judgment |
| 2. Holder and permission | Records holder role and governing agreement | Agreement clause, counsel note | Agreement limits use to services for the client |
| 3. Sampling | Draws a random sample across dates, sites, payers and systems | Sampling frame, seed, counts per stratum | Any system or site cannot be sampled |
| 4. Field necessity | Maps every field to a workflow step | Field necessity table | A field has no step and the owner insists on keeping it |
| 5. Coded identifiers | Checks identifier fields and how each is handled | Identifier inventory with method per field | Identifiers remain with no documented method |
| 6. Free text and attachments | Reads every free-text field and attachment type in the sample | Finding log with record references | Clinical narrative or third-party identifiers recur |
| 7. Special categories | Looks for records from substance use disorder programs and other separately protected sources | Source program list, flag counts | Any such record appears in the sample |
| 8. De-identification route | Names the method and who applies it | Method memo, expert report if used | No method is chosen or the method is unvalidated |
| 9. Sign-off | Records decision and conditions | Reviewer, date, version, conditions | Any earlier stop is unresolved |
Which identifiers hide in administrative fields and free text?
More than reviewers expect. Under the HIPAA Safe Harbor method described by HHS, a covered entity removes listed identifiers of the individual and of relatives, employers or household members, including names, geographic subdivisions smaller than a state, elements of dates other than year, telephone and fax numbers, email addresses, Social Security numbers, medical record numbers, health plan beneficiary numbers and account numbers, and must also have no actual knowledge that the remaining information could identify the person. The alternative route is a determination by a qualified expert that the risk of identification is very small.
Administrative data is full of these elements in places that look harmless: claim numbers that embed a member identifier, remittance trace numbers, payer correspondence addresses, appeal letters signed by a family member, and call notes that mention an employer. HHS also describes a limited data set, which removes direct identifiers but may keep dates and some geography; it is still protected health information, and its use for research, public health or health care operations requires a data use agreement. A reviewer should never describe a limited data set as de-identified.
- Claim and account numbers with embedded member or patient identifiers.
- Service dates, admission and discharge dates and birth dates in eligibility fields.
- Free-text denial reasons, appeal letters and call notes naming relatives or employers.
- Addresses and ZIP codes in billing and correspondence fields.
- Attachment file names that include patient names or record numbers.
Worked example (illustrative): a claim status sample with one record that should not be there
This example is constructed for this page and uses illustrative counts. A billing service proposes a claim status follow-up dataset and draws a stratified random sample of 400 claims across three payers and 24 months. The task statement is to predict the next follow-up action from status and adjustment codes.
The field necessity table drops diagnosis code fields, because no step of the status workflow uses them, and keeps service dates in generalized form. The identifier inventory finds that the claim number format embeds the patient account number, so the claim number is replaced with a keyed token that preserves joins across remittances. The free-text log finds 31 call notes, about 8% of the sample, that name a spouse or employer, so the notes column is excluded from the first release rather than scrubbed under time pressure.
The exception: 2 records come from a client that operates a substance use disorder treatment program. HHS's fact sheet on the 2024 Part 2 final rule describes separate federal confidentiality rules for these records, with their own consent requirements, including separate consent for SUD counseling notes. The reviewer stops, removes that client's records from the frame and sends the question to counsel. The limit of the recommendation is plain: the protocol finds the issue; it does not decide whether those records could ever be used.
Which findings end the review, and which only narrow the extract?
Findings about permission and special categories end the review until counsel resolves them; findings about fields and free text usually narrow the extract. Write the rule down before reviewing, so the outcome is not negotiated record by record.
| Finding | Effect on the review | Who resolves it |
|---|---|---|
| Agreement limits use to services for the client | Review stops | Counsel and the client relationship owner |
| Records from a separately protected program | Review stops for those records; frame narrowed | Counsel and privacy officer |
| Clinical narrative recurs in a free-text field | Field excluded or task narrowed | Privacy reviewer and data owner |
| Identifier embedded in a key field | Field tokenized before extract | Data engineer, checked by privacy reviewer |
| Field with no workflow step | Field dropped | Data owner |
| No de-identification method chosen | Extract postponed | Privacy officer, with an expert if expert determination is used |
How do payer versions and transaction formats affect what the sample shows?
They decide whether a sample from one period represents the rest. Payers change status code usage, attachment requirements and portal workflows; if the sample covers only the newest period, the reviewer has not seen the fields and free text that older records carry. Record the payer, transaction version and source system for each stratum, and add a stratum whenever any of them changed during the date range.
CMS's table of adopted standards gives version 5010 for the X12 claim, eligibility, remittance and prior authorization transactions, with compliance dates in 2012 and later operating rules for some of them. Proprietary portals and payer-specific attachments sit outside those standards, so the sample should include at least some records that came through each channel the full extract will contain.
Bringing a reviewed administrative sample to a licensing conversation
A completed protocol is what makes a licensing conversation short: it shows the task, the permission basis, the fields kept and the records excluded. For a revenue cycle company or billing service, SourceX can take that reviewed scope through rights review, contracting, preparation and delivery; the first step is a brief assessment that needs no records, and nothing is delivered until the company has approved the scope and terms and signed.
Frequently asked questions
How large should an administrative data sample be for a privacy review?
Large enough to include every source system, payer, site and period the full extract will contain, drawn at random within each. A few hundred records across those strata is a common starting point, but the strata matter more than the total. Record the sampling frame and seed so the draw can be repeated.
Is a limited data set the same as de-identified data?
No. HHS describes a limited data set as protected health information with direct identifiers removed that may still include dates and some geography. Its use for research, public health or health care operations requires a data use agreement, and it is not de-identified under HIPAA.
Can a business associate license administrative records it processes for clients?
Only if its agreements and the applicable rules allow it, which counsel must confirm. HHS describes business associates as performing functions on behalf of a covered entity under a business associate agreement that limits their uses. Processing the records does not by itself give a right to license them.
What should happen when free-text fields contain clinical detail?
Exclude the field from the first release or narrow the task so it does not need the field. Scrubbing clinical detail out of free text under time pressure is error-prone, and the administrative task rarely depends on it. Record the decision in the finding log.
Who should sign off on a healthcare administrative sample review?
A healthcare administrative expert should confirm the task and field necessity, and a qualified privacy reviewer should confirm permissions, identifier handling and special categories. Counsel decides any question about agreements or specially protected records.
Sources
- HHS guidance says covered entities must make reasonable efforts to use, disclose and request only the minimum protected health information needed for the purpose, and that routine or recurring disclosures may follow standard protocols. Source
- HHS de-identification guidance describes the Safe Harbor method (removing listed identifiers of the individual and of relatives, employers or household members, with no actual knowledge of identifiability) and the expert determination method. Source
- HHS describes a limited data set as protected health information with specified direct identifiers removed, usable for research, public health or health care operations only under a data use agreement. Source
- HHS defines business associates and lists functions such as claims processing or administration, data analysis, billing and practice management, performed under a business associate agreement. Source
- HHS's fact sheet on the 2024 42 CFR Part 2 final rule describes the confidentiality rules for substance use disorder patient records, including separate consent for SUD counseling notes. Source
- CMS lists the HIPAA adopted standards, including ASC X12N 837, 270/271, 835 and 278 version 5010, with their compliance dates and operating rules. Source
See if your company qualifies
A short company assessment. No data uploads are needed.