Skip to content

Systems and records

Prior authorization status dataset: keep requested, pending and approved apart

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Training data should record requested, pending and approved as separate, timestamped events on each service line rather than as one overwritten status. Requested is the office's submission, pending is a payer response deferring the decision, and approved, modified, denied or cancelled end the line's review. Keep raw review action codes, channel, urgency and payer category, validate the transitions, and leave clinical content out.

Provenance trail

Prior authorization status

This page covers prior authorization status dataset: keep requested, pending and approved apart.Illustrative

Key takeaways

  • A status column shows only the last answer; an event history shows the path and the waiting time.
  • Track the timeline per service line, because one request can end with different outcomes on each line.
  • Pending is a state of the case, not an outcome; a line that is still pending at extract time should be flagged open.
  • Keep the raw review action code beside the mapped event type, and publish the mapping.
  • Administrative timelines need dates, channel, actors and outcomes, not diagnoses or notes.

How should training data distinguish requested, pending and approved authorization?

Store requested, pending and approved as separate, timestamped events on each service line, never as successive values of one status column. Requested is an event the provider office creates when it submits; pending is a payer response that defers the decision; approved, modified, denied and cancelled are outcomes that end a line's review. A dataset that keeps those apart lets a model learn the sequence and the waiting time, not only the last answer.

The HL7 Da Vinci Prior Authorization Support implementation guide, which maps FHIR exchanges to the X12 278 transaction, makes the same distinction. It says whether a response is pended, denied or successful depends on the review action code on each item, and its examples use X12 review action codes such as A1 Certified in total, A3 Not Certified, A4 Pending and A6 Modified. It defines a pended item as one where the payer needs more time to make a final determination, so pending is a state of the case, not an outcome.

The guide also requires each returned item to echo the item sequence from the request, which it calls the main tracing identifier across requests and responses. That is the strongest hint for dataset design: the timeline lives at the item level. One authorization request with three service lines can end with one approved, one modified and one denied.

What does a single overwritten status field lose?

Most practice management and billing systems show the current status of a request, and a plain export captures only that value. The case then looks like approved, and everything that happened on the way disappears.

What disappears is exactly what an administrative model needs: how long the payer took, how many times the request was pended, whether the office had to send more information, whether the approval covered what was asked or a modified version, and whether the request was updated or cancelled midway. The Da Vinci guide's suggested metric data model records an initial submission time and a final response time per item, the number of responses including pended ones, and a final result of approved, modified, denied, pended or cancelled. Those fields only exist if the timeline was kept.

Timing also has a regulatory meaning now. CMS's Interoperability and Prior Authorization final rule, CMS-0057-F, requires impacted payers, other than qualified health plan issuers on the federally facilitated exchanges, to send decisions within 72 hours for expedited requests and seven calendar days for standard requests, generally beginning in 2026, and to give a specific reason for denials. A dataset without request and decision times cannot represent whether a case fell inside or outside those windows.

Which fields does the administrative state transition schema record?

The schema below records one row per event per service line, so the full history of a case is the ordered set of its rows. Every field is administrative; none carries a diagnosis, clinical narrative or attachment content. Field names are illustrative and should be versioned with the dataset.

Which fields does the administrative state transition schema record?
FieldType and allowed valuesPurposePrivacy treatment
case_keyOpaque token, stable within the releaseGroups the events of one authorization requestGenerated per release; not the payer's authorization number
item_seqInteger, echoed from the request lineTies every response to the line it answersNone needed
event_seq and event_timeInteger order and timestampOrders events and measures intervalsShift all times in a case by one offset to keep intervals
event_typesubmitted, updated, pended, info_requested, info_sent, approved, modified, denied, cancelledThe state change itselfNone needed
actor_roleprovider_office, payer, intermediaryShows who moved the caseRole only, never a person's name
channelx12_278, fhir_pas, portal, fax, phoneChannel affects timing and failure modesNone needed
urgencyexpedited or standardSelects the applicable decision windowNone needed
payer_categorymedicare_advantage, medicaid, chip, exchange_qhp, other_commercialDecides whether federal timeframes applyCategory only; payer names tokenized if kept
review_action_codeSource code as received, for example A1, A3, A4, A6Keeps the raw label behind event_typeNone needed
reason_categoryControlled list such as missing_information, not_covered, out_of_network, otherAdministrative reason without clinical narrativeFree-text reasons excluded

Which transitions are valid, and what should a validator reject?

Treat the history as a small state machine and reject any line whose events break it. A validator is cheap to write and catches most export errors before a reviewer ever reads a record.

The transition table lists what may follow what. Anything not listed is invalid for that line.

  • Times never decrease within a case after the shared offset is applied.
  • Every review_action_code maps to exactly one event_type in the published mapping table.
  • Lines still open at the extract cutoff carry an explicit open_at_cutoff flag rather than an invented outcome.
Which transitions are valid, and what should a validator reject?
FromAllowed next eventsValidator ruleTypical cause when it fails
(start)submittedThe first event on every line is submittedExport began after submission or merged two cases
submitted or updatedpended, info_requested, approved, modified, denied, cancelledA payer or intermediary event follows within the caseMissing payer response rows
pendedinfo_requested, approved, modified, denied, cancelled, updatedA pended line must later reach an outcome or be marked open at the extract cutoffTruncated history or extract taken mid-review
info_requestedinfo_sent, cancelledOnly the provider office sends informationActor roles swapped in mapping
info_sentpended, approved, modified, deniedA payer event followsDocumentation round trip not exported
approved, modified or deniedupdated, cancelledOutcomes are final for that version of the line; changes arrive as updatesStatus overwritten in place
modified(same as above)A modified line carries the authorized details as a linked added itemAuthorized quantity lost in export

Why does the schema stop at administrative facts and leave clinical detail out?

Keep what explains the administrative path and leave out what explains the clinical decision. The task here is learning how authorization requests move, not judging whether care was appropriate, so the record needs codes at a family level, dates, channel, actors, outcome and an administrative reason category.

The Da Vinci guide's privacy page notes that sharing information for prior authorization is subject to HIPAA's minimum necessary rule, and HHS describes that standard as limiting uses and disclosures of protected health information to the minimum necessary to accomplish the purpose. A training dataset is a different purpose from the original exchange, so its own pathway matters: for health data, HHS describes de-identification by safe harbor or by expert determination, and counsel confirms which route and permissions apply.

In practice, that means these items stay outside the schema.

  • Diagnosis codes and descriptions, clinical notes, attachments and imaging.
  • Free-text denial rationale and peer-to-peer call notes, which tend to restate clinical facts.
  • Patient demographics beyond what the task needs; usually none are needed.
  • Member IDs, authorization numbers, provider names and NPIs, replaced by release-scoped tokens or dropped.
  • Exact dates of service; keep intervals through a per-case offset instead.

What do a valid case history and a failing one look like?

Both histories below are illustrative and invented to show structure; they do not describe real cases or any SourceX dataset. Times are shown as hours after submission because the per-case offset hides calendar dates.

The valid history is one standard-urgency request through a payer portal with two service lines. Line 1: submitted at hour 0, pended at hour 20, info_requested at hour 20, info_sent at hour 31, approved at hour 70. Line 2: submitted at hour 0, modified at hour 70 with a linked added item recording the authorized quantity. The case closes when both lines reach an outcome, and the elapsed time for each line is computable.

The failing history comes from a status-only export. Line 1 has a single row: approved, time unknown. Line 2 has two rows: approved at hour 70 and pended at hour 90. The validator rejects line 1 because the first event is not submitted, and rejects line 2 because a pended event follows an outcome without an update, which usually means the status was overwritten and a later re-review was mislabeled. The fix is to re-export from the event or audit log, not to patch the rows.

How do payer rules and standard versions change the labels over time?

Labels drift when rules and standards change, so each row should carry enough context to interpret it. Record the payer category and urgency so that the CMS-0057-F decision windows are applied only where they apply, and record an era flag for decisions made before and after the rule's operational provisions took effect.

Record the source standard and version too. The Da Vinci guide is published in versions, and this page relied on its STU 2.2.1 source; review action codes come from an X12 code list, so keep the raw code as received alongside the mapped event_type and publish the mapping table with the dataset. When a payer or clearinghouse changes how it reports a modified authorization, the raw code is what lets a reviewer see the change instead of silently mislabeling it.

How do payer rules and standard versions change the labels over time?
Context fieldWhy it mattersExample values
rule_eraSeparates decisions made under different decision-window rulespre_2026, post_2026
source_standardChannel and standard change how statuses are reportedx12_278, fhir_pas_stu2
mapping_versionTies event_type to the raw code mapping usedmap_v3
extract_cutoffMarks lines still open when the data was pulledOffset hour of the extract

Where does the event history actually live in office systems?

Rarely in one place. The current status sits in the practice management or authorization work queue, while the events that make up the timeline are spread across audit logs, clearinghouse response files, portal histories and fax records. Rebuilding the history means joining these on the case and line, then checking the result with the validator.

Map each source before extracting, and note the gap each one leaves. The table is a starting inventory; the actual systems differ by office and vendor, so treat the rows as illustrative categories rather than product features.

Where does the event history actually live in office systems?
Source categoryEvents it usually holdsGap to watchHow to close it
Authorization work queueCurrent status, assigned staff, follow-up datesOverwrites earlier statusesPair it with the audit or change log
System audit or change logEvery status change with a time and userUser names and free-text commentsKeep the change and time; drop user and comment
Clearinghouse 278 response filesPayer responses with review action codesOnly covers electronic submissionsFlag channel so portal and fax cases are not judged incomplete
Payer portal status historyPortal-side submission and decision timesOften exported as screenshots or PDFsRecord only the event times and outcome
Fax and phone logsSent and received times for documentationPages contain clinical contentLog the event, never the page

What does an authorization history need before anyone licenses it?

Before an authorization history is offered for AI training, three questions are settled separately: who holds the records and under what agreements, whether privacy rules permit the new use or require de-identification first, and whether the extract keeps only the administrative fields in the schema. A billing company or practice may hold the data on behalf of clients, which makes the client agreements part of the rights review.

Volume and history depth matter for whether the work is useful. Depth matters too: AI buyers typically ask for two or more years of consistent history and tens of thousands of records or more in a single system, although a rare workflow can qualify with fewer and every case is decided on its own.

SourceX runs that sequence of sourcing, rights review, contracting and preparation for companies that license operational records to AI developers. A practice or billing company finds out whether it qualifies through a brief questionnaire with nothing uploaded, so it can learn whether its authorization history is a candidate before anyone exports a single event.

Frequently asked questions

What is the difference between pending and requested in prior authorization data?

Requested is the provider office's submission event, which starts the timeline for a service line. Pending is a payer response saying the decision needs more time or more information. Keeping them as separate events lets you measure how long payers took and how often requests were pended.

Is a modified authorization an approval?

It is an outcome, but not the same as approval in full. The Da Vinci guide returns the requested item with a Modified code and the authorized details as a linked added item. Store both so the dataset shows what was asked and what was authorized.

Why record review action codes if we already have event types?

The raw code is the evidence behind the mapped label. Payers and intermediaries do not always report outcomes the same way, and keeping the original code with a published mapping table lets reviewers audit or change the mapping later.

Do we need clinical notes to train an authorization workflow model?

Not for an administrative timeline task. The sequence, timing, channel, actors, outcome and an administrative reason category describe how requests move; clinical content belongs to a different task with its own review and permissions.

How do we keep intervals without exposing dates?

Apply one random offset to every timestamp within a case. Intervals and order stay intact, so time-to-decision remains measurable, while calendar dates are hidden. Record the method in the dataset documentation and have a privacy reviewer confirm it fits the release.

Sources

  • The Da Vinci PAS guide says whether a response is pended, denied or successful depends on each item's review action code, defines pended items as needing more time for a final determination, requires each item to echo its request sequence, and returns modified items with code A6 and authorized details as added items. Source
  • The Da Vinci PAS metric data model records per-item initial submission and final response times, the number of responses including pended ones, and a final result of approved, modified, denied, pended or cancelled. Source
  • The Da Vinci PAS privacy page says sharing information for prior authorization is subject to HIPAA's minimum necessary requirements. Source
  • CMS-0057-F requires impacted payers other than QHP issuers on the FFEs to send prior authorization decisions within 72 hours for expedited and seven calendar days for standard requests, generally beginning in 2026, and to give a specific reason for denials. Source
  • HHS describes the minimum necessary standard as limiting uses and disclosures of protected health information to the minimum necessary to accomplish the intended purpose. Source
  • HHS describes two de-identification methods for health information: safe harbor and expert determination. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies