Privacy and preparation
Healthcare scheduling exception dataset without clinical notes
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Scheduling exceptions can test an agent without clinical notes if each case is built from structured administrative fields: slots, status changes, cancellation codes, resource constraints and referral or authorization flags. Leave out the visit reason and free-text notes, review every remaining field for necessity, convert appointment dates to relative times, and grade the agent only on administrative outcomes such as slot, status and routing.
Evaluation split
Healthcare scheduling
Key takeaways
- Draw the nonclinical boundary field by field; FHIR itself calls the appointment reason more clinical than administrative.
- A signed task inventory lists each exception, its evidence, the fields excluded, the privacy question and the passing output.
- Store status histories, not snapshots; the exception is a transition and the test asks what happens next.
- Appointment dates and specialty or clinic names can identify or reveal conditions; transform or review them.
- Route any case whose right answer depends on clinical urgency to a person instead of grading the agent on it.
- De-identification does not create a right to license records a practice does not control.
How can scheduling exceptions test an agent without exposing clinical notes?
Build each test case from the structured administrative fields of the scheduling system, such as slots, appointment status changes, cancellation reason codes, resource and room constraints, and referral or authorization status flags, and leave out free-text notes and the clinical reason for the visit. Then review every remaining field for necessity, de-identify the set, and grade the agent only on administrative outcomes: the right slot, the right status, the right routing.
Scheduling exceptions are a good fit for this because the hard part of the work is administrative. A provider calls in sick and 18 appointments need new slots. A patient no-shows and the clinic's rebooking policy applies. A specialist visit cannot be booked because the referral has not arrived. None of these decisions needs the visit's diagnosis or a clinician's note; they need calendars, rules and status histories.
This page is written for practice administrators, revenue cycle and patient access leaders, and the privacy officers who review their data. It is about administrative records only and is not clinical guidance. Nothing here should be used to decide clinical urgency or care.
Where the nonclinical boundary runs through an appointment record
The boundary runs through individual fields, not whole tables, so decide field by field. The HL7 FHIR Appointment resource is a useful map because its own definitions separate the two kinds of content. It describes Appointment.reason as the reason the appointment is being scheduled and says this is more clinical than administrative; it describes Appointment.note as additional comments for internal use; and it separates patientInstruction, which carries patient-facing instructions such as bringing a referral.
Most practice management and scheduling systems hold equivalents of these fields even if they do not use FHIR. The table below sorts the common ones. Keep means the field is usually needed for an administrative scheduling task; transform means keep a coarser version; exclude means it stays out of the test set.
Check custom fields as carefully as standard ones. Practices often add free-text or picklist fields over the years, such as a scheduler comment, a visit purpose dropdown or a prep checklist, and these tend to collect clinical detail even when their label sounds administrative. Profile a sample of values for every custom field before deciding, and treat any field with unpredictable free text as excluded until a reviewer has read a sample.
| Field (FHIR name where one exists) | Typical content | Decision | Reason |
|---|---|---|---|
| status | Booked, canceled, no-show, waitlist and similar | Keep | The exception is defined by status changes |
| cancellationReason | Coded reason such as patient or provider canceled | Keep coded value only | Drives rebooking and policy rules |
| start, end, minutesDuration | Appointment times and length | Transform to relative times | Needed for slot logic; exact dates are identifiers |
| slot and participant (practitioner, location, device) | Which provider, room or equipment was booked | Keep as tokens | Resource conflicts are the core of many exceptions |
| appointmentType | Style of visit, such as follow-up or walk-in | Keep | Determines slot length and template rules |
| serviceType and specialty | Kind of service and specialty | Transform or review | Some specialties reveal sensitive conditions |
| reason | Why the visit is scheduled | Exclude | FHIR itself calls it more clinical than administrative |
| note | Internal free-text comments | Exclude | Staff often type symptoms or history into notes |
| patientInstruction | Patient-facing prep instructions | Exclude by default | Can reveal procedures and conditions |
The privacy reviewed administrative task inventory
List every scheduling exception task the agent will be tested on, and for each one record the administrative evidence it needs, the fields deliberately left out, the privacy question a reviewer answered and what a passing agent output looks like. The inventory is the asset a privacy reviewer signs; tasks not in it are out of scope. Rows are illustrative and should be adapted to the practice's own workflows.
At a glance
The privacy reviewed administrative task inventory
Start the inventory from the exceptions staff actually handle most often, which the scheduling system can usually report by cancellation reason code and reschedule count. A short, well-reviewed inventory of six to ten tasks is more useful than a long one whose privacy questions were answered in a hurry, and each row added later goes through the same review before any cases for it are built.
| Exception task | Administrative evidence needed | Excluded on purpose | Privacy review question | Passing agent output |
|---|---|---|---|---|
| Provider absence, bulk reschedule | Affected slots, provider templates, open slots for equivalent providers | Visit reason, notes | Can equivalence be judged from appointment type alone? | Each affected visit offered a compliant slot or routed to staff |
| Double-booked slot | Slot, both bookings, booking timestamps, template rules | Notes, instructions | Are booking user IDs needed, or only roles? | Correct booking kept per rule; other moved or routed |
| No-show rebooking | No-show status, practice no-show policy version, prior no-show count | Reason, notes, diagnosis codes | Is the prior count needed or only a policy threshold flag? | Policy applied as written; exceptions routed |
| Waitlist backfill | Canceled slot, waitlist entries, appointment type, requested period | Reason, specialty if sensitive | Does waitlist priority depend on clinical urgency? | Offer order follows the administrative priority rule |
| Referral missing before specialist booking | Referral received flag and date offset, required-referral rule | Referral letter content | Can the flag replace the referral document? | Booking held and referral request routed |
| Authorization pending | Authorization status flag, payer rule version | Clinical justification, attachments | Is payer identity needed, or only the rule version? | Visit not confirmed until status allows; routed |
| Interpreter or accessibility need | Need flag, interpreter resource availability | Free-text detail | Is a yes or no flag enough? | Booked with resource or routed |
| Room or equipment conflict | Location and device bookings, maintenance blocks | Procedure detail | Does device type reveal a condition? | Conflict resolved or escalated |
| Entered in error | Status history showing entered-in-error | Everything else | None beyond status | Record excluded from rebooking logic |
Which status changes turn a booking log into a test case?
An exception is a transition, so store status histories, not snapshots. A booking log that keeps only the current status shows that an appointment was canceled but not that it was booked, moved twice and then canceled by the clinic. The agent is tested on what it should do next at each step, which only a history can show.
The FHIR appointment status code system gives a vocabulary most systems can map to: proposed, pending, booked, arrived, fulfilled, cancelled, noshow, entered-in-error, checked-in, waitlist and unknown. FHIR defines checked-in as the point when all pre-encounter administrative work is complete and the encounter may begin, and waitlist as an appointment placed on a waitlist to be scheduled when a slot is available. Mapping local codes to these values, and keeping the local code beside the mapped one, makes test cases comparable across practices.
Each test case then has three parts: the history up to the exception, the event that created it (a provider block, a cancellation, a missing flag), and the expected administrative outcome. Keep the version of any policy the outcome depends on, such as the no-show rule or the referral requirement, because the right answer changes when the policy does.
How should scheduling test cases be graded and split?
Grade the agent on whether each action is permitted and complete, not on whether it picked the same slot a scheduler picked. In most exceptions several slots satisfy the template and the policy, so a scoring rule that demands the historical slot punishes correct answers. Check four things instead: the proposed slot exists and fits the appointment type, the policy version in force allows the action, every affected appointment received an action or a routing, and anything the agent could not decide from administrative fields went to a person.
Split the cases by provider, location and time period rather than at random. Scheduling behavior is highly local: templates, staff habits and payer mixes differ between clinics, and a random split lets the agent memorize one clinic's patterns and look better than it is. Hold out at least one location and one recent period entirely, and keep a separate set of cases where the correct outcome is routing, because an agent that never routes will otherwise score well on the easy majority.
Record the grading rule beside each case in the inventory, so the privacy reviewer can see that no grading step depends on an excluded field.
Reviewing each field against the minimum necessary standard
Use the minimum necessary standard as a field-by-field filter even when the data will later be de-identified. HHS guidance describes the standard as requiring reasonable efforts to limit uses, disclosures and requests of protected health information to the minimum necessary for the intended purpose, and says covered entities make their own assessment of what is reasonably necessary. The inventory is the written record of that assessment for this purpose.
Scheduling itself is ordinary administrative work for a practice. HHS describes health care operations as certain administrative, financial, legal and quality improvement activities a covered entity needs to run its business and support treatment and payment. Preparing a dataset for an outside AI developer is a different purpose, so the review must ask again, for that purpose, which fields are needed. This is general information, not legal advice; the practice's privacy officer and counsel decide the legal basis for any use or disclosure.
- Is the field needed to decide the administrative outcome in at least one inventory task? If not, drop it.
- Can a coarser version serve the task: a flag instead of a document, a role instead of a user ID, a relative time instead of a date?
- Could the field reveal a condition by itself or in combination, such as a specialty, a device type or a location name?
- Who reviewed it, on what date, and against which version of the inventory?
Dates, clinic names and other quiet identifiers in schedules
Appointment dates are identifiers even when names are gone. HHS's de-identification guidance explains that the Safe Harbor method removes all elements of dates, except the year, that are directly related to an individual, and gives admission and discharge dates as examples. Appointment start times sit squarely in that category, which is why the field table above converts them to relative offsets from an anchor event.
Relative offsets keep what the agent needs: the gap between cancellation and the new slot, the day of week, the time of day and the order of events. Whether a given transformation satisfies Safe Harbor, or whether the set needs the Expert Determination method, where a qualified expert finds the risk of identification very small, is a decision for the privacy reviewer. HHS describes both methods; this inventory does not choose between them.
Clinic, department and specialty names can be more revealing than dates. A booking at an oncology infusion chair, a behavioral health clinic or a substance use disorder program discloses a condition without a single diagnosis code. Records of federally assisted substance use disorder programs are also governed by 42 CFR Part 2, which HHS updated in a 2024 final rule, and they should be left out of a general scheduling inventory unless counsel has reviewed them specifically.
Worked example: an illustrative provider-absence reschedule batch
Consider an illustrative outpatient practice where one provider's clinic day is blocked at short notice. All counts are invented. The schedule holds 18 booked appointments for that provider; the test case gives the agent the slot list, the appointment types, the templates of three other providers and the practice's reschedule policy version.
The expected outcome, built from what staff actually did, is that 12 appointments move to equivalent slots the same week, 4 are offered later slots because the appointment type requires that specific provider, and 2 are routed to staff. The agent is graded on whether each of its proposals is permitted by the template and policy, not on matching the human's choice of slot.
The exception is one of the two routed appointments. Its appointment type is ordinary, but the internal note, excluded from the test set, said the patient needed same-week follow-up. Staff knew that; the agent cannot. The correct test expectation is therefore not a specific slot but a routing: when an appointment carries a staff-set priority flag the agent cannot see the basis for, it hands the case to a person. Building that expectation into the inventory keeps the test administrative and stops the agent from being rewarded for guessing at clinical urgency.
Scheduling tasks that do not belong in an administrative test set
Leave out any task whose correct answer depends on clinical judgment, and stop to request specialist review when a task cannot be specified without clinical fields. Triage by symptoms, deciding how urgent a visit is, choosing which test to schedule and deciding whether a patient can wait are clinical decisions, even when a scheduler carries them out under a protocol.
Also pause when the privacy review cannot make a field coarse enough to be safe, when the practice's agreements with payers or health systems restrict reuse of scheduling data, or when the records come from a vendor platform whose contract limits the practice's rights. De-identification addresses identifiability; it does not give a practice a right to license records it does not control.
- Symptom-based triage or urgency scoring: exclude.
- Tasks that need the visit reason or clinical notes to grade: exclude or redesign as a routing task.
- Behavioral health, reproductive health or substance use disorder program schedules: exclude pending specific review.
- Fields that cannot be coarsened below the reviewer's risk threshold: drop the task, not the threshold.
Licensing a reviewed scheduling exception set through SourceX
Medical groups, billing companies and practice management operators that hold years of scheduling histories can license prepared, de-identified copies to AI developers through SourceX, which runs qualification, rights review, contracting, preparation, delivery and payment. Scope, including the inventory of tasks and fields, is approved and the agreement signed by the company before any records are delivered.
SourceX performs de-identification through approved processing routes, and the field decisions in this inventory are the natural input to that work. A short assessment, with no data uploaded, shows whether a company is a fit.
Frequently asked questions
What is a healthcare scheduling exception dataset?
It is a set of appointment histories where something went off plan, such as a provider absence, a no-show, a double booking or a missing referral, paired with the administrative action staff took. It is built from structured scheduling fields and status changes, not from clinical documentation. It is used to test or train agents that handle rescheduling and routing.
Which scheduling fields are clinical and should stay out?
The reason for the visit, internal free-text notes and most patient instructions should stay out, because they often contain diagnoses, symptoms or procedures. The FHIR Appointment definition itself describes the reason field as more clinical than administrative. Specialty and location names need review because some reveal conditions on their own.
Are appointment dates protected health information?
When linked to an individual, appointment dates are dates related to the provision of care, and HHS's Safe Harbor method removes date elements other than the year. Converting dates to relative offsets keeps the sequencing an agent needs, but the privacy reviewer decides which de-identification method the set relies on.
Can an AI agent decide which patients need urgent appointments?
Not in an administrative test set. Urgency depends on clinical judgment, so any appointment with a staff-set priority the agent cannot see the basis for should be routed to a person. Grading an agent on clinical urgency without clinical fields would reward guessing.
Who should review a scheduling dataset before it is shared?
A healthcare administrative expert to confirm the tasks and expected outcomes, and a qualified privacy reviewer to confirm field necessity and de-identification. Counsel confirms the legal basis and any contractual limits on reuse of the records.
Sources
- The FHIR Appointment definition describes the reason element as more clinical than administrative and separates internal notes from patient instructions. Source
- The FHIR appointment status code system defines statuses including booked, cancelled, noshow, checked-in, waitlist and entered-in-error. Source
- HHS guidance says covered entities must make reasonable efforts to limit uses, disclosures and requests of PHI to the minimum necessary and make their own assessment of what is needed. Source
- HHS describes health care operations as administrative, financial, legal and quality improvement activities needed to run the business and support treatment and payment. Source
- HHS de-identification guidance explains Safe Harbor's removal of date elements other than the year and the Expert Determination standard of very small risk. Source
- HHS issued a 2024 final rule updating 42 CFR Part 2, which governs substance use disorder program records. Source
See if your company qualifies
A short company assessment. No data uploads are needed.