AI uses for records
Appointment scheduling constraint dataset: five answers when an MRI moves
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
An appointment scheduling constraint dataset for outpatient imaging should label what the scheduler decided when a booked MRI or CT had to move: rebook, hold a slot on a named condition, substitute a scanner or site, waitlist, or escalate. Each case freezes the constraints visible at that moment, keeps coded flags instead of clinical detail, and grades the decision by whether the exam happened inside the ordered window.
Agent run
Appointment scheduling
Key takeaways
- The decision has five answers, and three of them are not a new time: a conditional hold, a waitlist entry or an escalation can each be correct.
- Freeze six record types as they stood when the scheduler acted: the order, the appointment history, the scanner and session schedule, the authorization log, readiness checks and authorized preferences.
- In HL7's 32 published cancellation reason codes, 10 name a readiness check that must clear before rebooking, and 7 state a fact about the patient's health; store the family, not the code.
- Every case needs a hard negative, an open slot that breaks a constraint recorded outside the calendar.
- Score the disposition before the slot, credit correct deferrals, and penalize unconditional bookings made while a check is pending.
- Licensed scheduling records show decisions and outcomes, not counterfactuals, and need privacy and rights review before any extraction.
What decision does an appointment scheduling constraint dataset capture?
It captures one decision: what a scheduler does when a booked outpatient MRI or CT exam has to move and the obvious open slot breaks at least one constraint. In an imaging center the right answer is often not a new time. It can be a slot held until a check clears, another scanner or site, a place on a waitlist, or a handoff to the ordering clinician, and a dataset that records only rebooked times teaches a model to force a booking where an experienced scheduler would wait.
Five dispositions cover the decision. Rebook confirms a replacement slot that meets every hard constraint. Conditional hold reserves a slot pending a named check, with a date on which the hold is released. Substitute moves the exam to an equivalent scanner, site or session. Waitlist keeps the current booking, or none, and records the period the patient would prefer. Escalate sends the case to the ordering clinician or the radiology team, because the only feasible dates fall outside the ordered window or a flag needs clinical judgment.
The standards already name the partial states. In the current FHIR build, an appointment is pending when some participants have not finalized their acceptance and waitlisted when it is to be scheduled or confirmed once a slot or service is available, and a participant can be tentative. FHIR's own notes describe a waitlist flow in which a patient keeps an inconvenient booking while a second appointment waits with a requested period, then accepts a better slot when one is offered. The Argonaut scheduling guide adds a hold operation that temporarily prevents a proposed appointment from being booked by another client while the booking is completed.
Which records does an imaging center keep for a reschedule?
The decision draws on six record types, usually split across the radiology scheduling module, the order or referral record, the payer authorization log and an outreach log. A case is useful only when all six are frozen as they stood when the scheduler acted, because a slot that opened an hour later did not exist for that decision.
Three details decide whether the history survives. FHIR's definition of the cancellation date says that when an appointment is cancelled more than once it reflects only the most recent cancellation, so the audit log, not the appointment row, holds the sequence. FHIR binds the cancellation reason to its value set only as an example, so every center's codes are local and need a published mapping. And FHIR's comment on the appointment note says it can carry guidance on when to book, including restrictions, which means constraints often sit in free text next to clinical remarks and have to be extracted into coded fields.
For each record type, extract the decision fields in coded form and leave the rest behind:
| Record | Fields the decision needs | Kept as | Left out |
|---|---|---|---|
| Order or referral | Exam type, needed-by window, priority, ordering clinician role | Exam family code, window as day offsets, priority tier | Indication, diagnosis and clinical history |
| Appointment history | Every status change with time and actor role, replaced-appointment links, cancellation reason | Event sequence in offsets; local reason mapped to a published family | Free-text notes and patient instructions |
| Scanner and session schedule | Slots by scanner capability, session staffing such as sedation or contrast support, maintenance blocks | Capability flags on every slot | Staff names |
| Authorization log | Status, submission and decision dates, approved date range | Status code and day offsets | Clinical justification and attachments |
| Readiness checks | Safety screening, preparation, results or clearance, each cleared, pending or do not proceed | One flag per check with its date offset | Screening answers and result values |
| Patient constraints and authorized preferences | Requested periods, declined sites, interpreter language, transport need, who recorded each and when | Coded constraints with source and date | Contact details and anything the patient did not authorize staff to store |
What HL7's 32 cancellation reason codes say about constraints
The published vocabulary shows how much of a reschedule is a constraint problem. HL7's appointment cancellation reason code system, version 2.1.1 in the HL7 terminology repository at commit 828b713, has 32 codes. Classified for this article by what a rescheduler must do next, 19 imply a replacement may be needed, 4 mean the appointment is no longer needed, 6 record a channel or an error, and 3 are top-level groupings.
The 19 fall into three families. Seven name a resource that failed: a provider's absence, equipment, a room, an interpreter or the weather. Two name a patient circumstance that changes which slot or site works, transport and a move of residence. Ten name a readiness check that has to clear before any new booking is safe: coverage, authorization, finances, preparation, results, screening or a clinical change. Those ten are where conditional holds and escalations come from.
How the count was made: every concept in the file counts as a code, top-level groupings included, and each receives one class from its display text. The classes are resource (prov-pers, prov-edu, prov-hosp, maint, oth-int, oth-room, oth-weath), readiness (pat-mt, pat-preg, prov-labs, prov-mri, prov-onc, meds-inc, oth-cms, oth-fin, oth-iv, oth-mu), logistics (pat-lt, pat-mv), ended (pat-dec, pat-fb, pat-ucp, prov-dch), process (pat-crs, pat-cpp, pat-swl, oth-err, oth-oerr, oth-swie) and grouping (pat, prov, other). The calls are editorial, and moving residence could as well be read as ending the need; the script that produced the counts, xs0802_cancel_codes.py, holds the full list, stops if the file and the list disagree, and rounds the one percentage once, half up, from an exact fraction.
Seven of the 32 codes, 21.9%, state a fact about the patient's health or vital status, such as a pregnancy, laboratory results out of range, an MRI screening form marked do not proceed or an oncology treatment change, and five of those seven are readiness codes; one more reports a provider's hospitalization. A dataset keeps the decision signal and drops the health fact by storing the family and the type of check, not the code:
| Family (codes) | What the rescheduler must do | Codes that state a health fact | Store in the dataset as |
|---|---|---|---|
| Resource failed (7) | Offer another slot, scanner, site or interpreter session | None about the patient; one about a provider | Resource family and the type of resource that failed |
| Readiness check pending (10) | Hold or escalate until the check clears | Five: pregnancy, laboratory results, MRI screening, oncology plan, IV access | Check type (coverage, authorization, preparation, clinical clearance) and its status |
| Patient logistics (2) | Offer a slot or site that solves transport or distance | None | Logistics family and the constraint it adds |
| Need ended (4) | Close the case with no replacement | Two: deceased, feeling better | Ended, without the reason text |
| Channel or error (6) and groupings (3) | Nothing about the next slot; read the history instead | None | A process flag, kept out of decision labels |
The label schema: five dispositions and how each outcome is known
Each case carries two labels: the disposition, judged on what was knowable at decision time, and the outcome, read later from the appointment history. Keeping them apart stops hindsight from grading the decision, since a correct hold can still end in a cancellation and a careless rebooking can get lucky.
Use two adjudicators with scheduling experience for every case, and add a radiology reviewer when a case carries a clinical flag. Record disagreements rather than forcing a consensus, and move cases the adjudicators could not settle out of the scored set into a separate review file, where they show which constraints the schema still describes badly.
Adjudicators record one primary disposition per case and, where two specialists agree it would also have been acceptable, one alternative from the same five. The outcome window comes from the order, not from the dataset builder: an exam is complete when it reaches arrived or fulfilled inside the ordered window with no later cancellation, and a case still open at extraction is censored rather than failed. The schema sets the evidence and the outcome reading for each disposition:
| Disposition | Correct when, at decision time | Evidence in the record | Outcome and how it is known |
|---|---|---|---|
| Rebook | A slot meets every hard constraint and every authorized preference inside the ordered window | New booking linked to the appointment it replaces | Exam arrived or fulfilled in the window with no further cancellation |
| Conditional hold | A feasible slot exists but a named check is still pending | Slot held as pending or tentative, with check type and release date | Check cleared and exam done in the window, or hold released on its date |
| Substitute | The original scanner, site or session cannot meet a constraint and an equivalent one can | Booking on another resource with matching capability flags | Exam done in the window on the substitute resource |
| Waitlist | No acceptable slot exists now, or the patient keeps a booking and asks for a better period | Waitlist entry with the requested period | Better slot accepted, original kept, or still waiting at extraction (censored) |
| Escalate | Only slots outside the ordered window exist, or a flag needs clinical judgment | Routing to the ordering clinician or radiology team with a trigger code | Documented response: new window, changed order or cancellation by the clinician |
A casebook of six reschedules and the tempting wrong answer
A casebook pairs each reschedule with the open slot that looks right and is not, and six cases are enough to cover all five dispositions at least once.
Cases 2 and 5 carry the most teaching value. In case 2 the tempting slot passes every calendar and resource check and fails only because the payer's decision is still due, so a model has to read the authorization log and the payer's timeframe to see the problem. In case 5 every constraint can be met, but not inside the order's window, which turns the best available slot into a question for the clinician rather than a booking.
The six cases are illustrative, written for this article, and describe no real center or patient. Each pairs the constraints visible at decision time with the open slot a naive scheduler or model would take, the hard negative, and the adjudicated disposition from the schema's five:
| Case (illustrative) | Constraints at decision time | Tempting slot (hard negative) | Adjudicated disposition |
|---|---|---|---|
| 1. Scanner down for repair | The booked MRI scanner is blocked for two days; the ordered window ends in ten days; the patient authorized weekday mornings at either of two sites | An afternoon slot on the same scanner after the repair | Substitute: a morning slot on the equivalent scanner at the second site |
| 2. Authorization pending | The payer received the request yesterday; the earliest open slot is tomorrow; the window ends in three weeks | Tomorrow's slot, booked unconditionally | Conditional hold: a slot after the payer's decision timeframe, released if the request is denied |
| 3. Interpreter unavailable | A recorded need for a language interpreter; the next slot with interpreter coverage is in six days | Tomorrow's slot without interpreter coverage | Rebook: the slot in six days with interpreter coverage |
| 4. Screening marked do not proceed | The safety screening flag is set; slots are open all week | Any open slot this week | Escalate to the radiology team; no booking until cleared |
| 5. Only late slots | Every slot in a sedation-staffed session falls after the ordered window | The first sedation-staffed slot, outside the window | Escalate to the ordering clinician with the earliest feasible date |
| 6. Patient asks for a better time | The patient keeps the booked slot but prefers evenings; evening sessions are full | Cancel and rebook into the first evening opening, five weeks out | Waitlist: keep the booking and record the requested period |
Hard negatives: open on the calendar, wrong in the record
A hard negative is an open slot that passes the calendar check and fails a constraint recorded somewhere else. Hard negatives separate a model that has learned constraints from one that has learned to fill gaps, so every case should carry at least one, labeled with the constraint it breaks.
Calendar search hides most constraints by design. The Argonaut scheduling guide limits its availability search to a few simple inputs, practitioner, available times, location, specialty and common visit types, and leaves the server to work out which schedulable assets a visit needs. Five kinds of hard negative recur in imaging as a result: a slot on a scanner or session without a required capability, such as sedation support; a slot before a pending check can clear; a slot that breaks a preference the patient authorized; a slot without the interpreter or transport the record says is needed; and a slot outside the ordered window, where the right move is escalation, not the earliest date.
Two published rules sharpen these kinds. CMS says its Interoperability and Prior Authorization final rule requires impacted payers, including Medicare Advantage plans and Medicaid and CHIP programs, to decide standard requests for medical items and services within seven calendar days and expedited requests within 72 hours from 2026, with the clock starting when the payer receives the request. An ACR-ASNR-SPR practice parameter for MR states: "Patients and all other persons entering the MRI safety zone must be screened and interviewed (if their condition permits) before the examination." It also notes that sedation may be needed to achieve a successful examination. A slot that ignores either is a hard negative, whatever the calendar shows.
How to score a model on rescheduling cases
Score the disposition first and the slot second, and reward a correct deferral as much as a correct booking. A model that always takes the earliest slot will look efficient on days to appointment and fail every case in which waiting was right.
At a glance
How to score a model on rescheduling cases
The rubric gives each case up to four points, and totals are reported by disposition, because a model can score well overall while never choosing to hold or escalate:
Checklist · 6 items
- Disposition, 2 points: the model's choice matches the primary label; 1 point when it matches the recorded alternative instead.
- Hard constraints, 1 point: any proposed slot meets every hard constraint in the frozen record. A slot that breaks one scores zero for the whole case, whatever the disposition.
- Condition named, half a point: a hold or an escalation names the check or trigger and the date the case will be revisited.
- Preference fit, half a point: any proposed slot respects every preference the patient authorized; preferences recorded without authorization are not scored. When the primary label proposes no slot and the model proposes none, both slot items are awarded.
- False certainty, minus 1 point: an unconditional booking while a readiness check is pending.
- Reporting: mean points per case for each of the five dispositions, plus the share of proposed slots that broke a hard constraint.
0 of 6 ticked. Tick each item as you go.
What an imaging center's records can and cannot show
They show what schedulers decided and what happened next; they cannot show what a different choice would have produced, and they hold only the constraints someone recorded. A center's records can support training and evaluation on the decision, provided the limits travel with the data.
Rights and privacy come first. Scheduling records about patients are health information, and HHS guidance on the minimum necessary standard says routine disclosures can follow standard protocols limited to the information needed for that type of disclosure, while using an entire medical record needs an explicit justification. A field map approved by the privacy officer, a documented de-identification method and counsel's view of who controls the records and may license them come before any extraction.
Four further limits apply to any licensed set:
- Counterfactuals are missing. When a scheduler held a slot, the record cannot say whether an immediate booking would have worked, so outcome labels grade only the path taken.
- Preferences are only as good as their recording. A preference the patient never stated, or staff never stored, is invisible, and a slot that looks acceptable may not have been.
- Local codes differ. Because FHIR's cancellation reason binding is an example, two centers' codes rarely match; the family mapping above is a starting point each source must review.
- Schedules change. Session templates, scanner capabilities and payer rules move over time, so each case should carry the version of each that was in force at decision time.
Requesting a rescheduling decision set
A team that needs these cases should name the decision, the five dispositions and the outcome window rule in its request, along with the acceptance test it will run on a sample. SourceX sources datasets for AI teams on request from U.S. businesses rather than from stock, so a request describes what the cases must contain and the format, timing and licensing terms needed; nothing here implies that imaging scheduling records are on hand today.
A useful request also states the share of cases that must carry a hard negative, the record types from the extraction table, and the privacy treatment, such as cancellation reasons stored as families. Any source company approves the scope and terms and signs its agreement before a single record is delivered.
Frequently asked questions
What is an appointment scheduling constraint dataset?
It is a set of scheduling cases, each frozen at the moment a booked appointment had to move, recording the constraints in force, the disposition chosen and the outcome read later from the appointment history. Its purpose is to teach or test whether a model respects constraints held outside the calendar.
Why include cases where the right answer is to wait?
Because a model trained only on completed rebookings learns to fill the earliest gap. Holds, waitlists and escalations are where constraint-aware scheduling differs from calendar filling, and a scoring rubric should reward them when they are correct.
Does a scheduling dataset need diagnoses or visit reasons?
No. The decision needs coded constraints: exam family, needed-by window, capability flags, check statuses and authorized preferences. Cancellation reason codes that state health facts can be stored as their family, such as readiness check pending, without the underlying reason.
How is the outcome of a reschedule known?
From the appointment history after the decision: whether the exam reached arrived or fulfilled inside the ordered window without a further cancellation. Cases still open at extraction are censored rather than counted as failures.
Can an imaging center license its scheduling records for AI?
Possibly, but only after counsel confirms who controls the records and on what terms, and a privacy officer approves the field map and the de-identification method. Scheduling records about patients are health information, so the minimum necessary standard and the organization's contracts shape what can leave.
Sources
- HL7's appointment cancellation reason code system, version 2.1.1, at commit 828b713: 32 codes in patient, provider and other groups plus equipment and preparation codes, classified for this article. Source
- FHIR's appointment notes describe typical status flows, including re-negotiation and a waitlist flow in which a patient keeps a booked appointment while a waitlisted one carries a requested period. Source
- The FHIR Appointment definition: cancellationReason bound to an example value set; cancellationDate reflects the most recent cancellation; note can carry guidance on when to book, including restrictions; participant status can be tentative. Source
- FHIR appointment status codes define pending and waitlist. Source
- The Argonaut scheduling guide limits availability search to simple inputs (practitioner, available times, location, specialty, common visit types) and defines a hold operation that temporarily prevents booking by another client. Source
- CMS says the Interoperability and Prior Authorization final rule requires impacted payers to decide standard prior authorization requests for medical items and services within seven calendar days and expedited ones within 72 hours from 2026, with the clock starting when the payer receives the request. Source
- An ACR-ASNR-SPR MR practice parameter states that patients and all other persons entering the MRI safety zone must be screened and interviewed, if their condition permits, before the examination, and that sedation may be needed. Source
- HHS guidance on the minimum necessary standard says routine disclosures can follow standard protocols limited to the minimum necessary for that type of disclosure, and use of an entire medical record needs an explicit justification. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.