Clinic no-show review for service leads

Nothing here is clinical advice, and nothing here should touch identifiable patient information.

Use one de-identified appointment extract and a small sample of de-identified contact-note themes to produce a decision brief each month. This is for outpatient operations managers and service leads who need to choose practical changes, rather than receive a long list of possible reasons.

This workflow examines administrative patterns only. It is not clinical advice, does not assess individual patients, and must not include identifiable patient information in the material you provide to the model.

1. Set the review question and period

Choose one clinic, service line, or appointment type for each review. Do not combine services with different booking routes, patient populations, or attendance rules. A mixed dataset produces vague findings.

Set a fixed review period, such as the last complete calendar month. Write one question in the review document:

Which administrative factors appear most associated with unattended booked appointments, and which two changes should we test next month?

Also record:

• The denominator, meaning all appointments booked in the period.
• Your local definition of no-show. Keep this separate from late cancellations and provider cancellations.
• The decision owner, usually the service lead.
• The changes already made during or shortly before the period.
• The operational constraint, such as no additional call capacity or no change to booking-system configuration.

2. Prepare a de-identified review pack

Create a spreadsheet or CSV with one row per appointment. Remove names, dates of birth, addresses, phone numbers, email addresses, medical record numbers, free-text notes, and any internal identifier that could be linked back to a person. Replace exact appointment dates with a week number or month where that still supports the analysis.

Include only fields needed to test an operational hypothesis. A useful minimum set is:

Field	Example values	Why it matters
Attendance outcome	attended, no-show, late cancellation	Defines the result being reviewed
Appointment type	new, follow-up, procedure	Shows whether patterns differ by visit type
Booking lead time band	0-2 days, 3-7 days, 8-21 days, 22+ days	Tests whether long waits are associated with non-attendance
Session band	morning, afternoon, evening	Identifies timing patterns without exact dates
Day of week	Monday to Friday	Identifies scheduling concentration
Booking route	portal, phone, staff-booked	Tests friction in the booking process
Reminder status	sent, not sent, unknown	Checks the delivery process, not a patient's behaviour
Reminder channel	text, email, phone, none	Compares available contact routes
Cancellation window	none, under 24 hours, 1-3 days, 4+ days	Separates recoverable slots from no-shows

Keep a second, separate note containing the field definitions and data quality issues. Do not upload this if it contains patient identifiers.

For contact notes, do not provide raw note text. Ask an authorised staff member to create a de-identified theme count first. For example: transport mentioned: 12, could not reach clinic: 9, work conflict mentioned: 7, reason not recorded: 31. These are reported reasons, not verified causes.

Check your organisation's information-governance process before using any external tool. Product capabilities and data-handling options can be version-dependent, so confirm the current position in the xAI documentation.

3. Calculate the baseline before asking for interpretation

Calculate the overall no-show rate yourself:

no-show rate = no-show appointments / booked appointments

Then calculate the same rate for each meaningful group, such as appointment type, lead-time band, reminder status, and session band. Include the count as well as the rate. A 25% rate from four appointments is not comparable with a 12% rate from 400 appointments.

Use this structure in your working sheet:

Group	Booked	No-shows	No-show rate	Difference from overall rate
Overall	[count]	[count]	[rate]	n/a
22+ day lead time	[count]	[count]	[rate]	[percentage points]
Reminder not sent	[count]	[count]	[rate]	[percentage points]

Do not ask the model to invent calculations from a narrative description. Give it the completed summary table and the de-identified theme counts.

4. Ask for operational hypotheses, not patient judgements

Paste the summary table, contact-note theme counts, definitions, and constraints into a new conversation. State that the material is de-identified and that the task is administrative analysis only.

Use a prompt like this:

You are helping prepare an outpatient operations decision brief. Analyse only the de-identified aggregate data below. Do not infer clinical, personal, or protected characteristics. Do not state that any factor causes no-shows.

1. Identify up to three attendance patterns supported by the figures.
2. For each pattern, distinguish the observed fact from possible operational explanations.
3. Flag missing data, small groups, and confounding factors that weaken each finding.
4. Propose up to three low-risk operational tests within these constraints: [insert constraints].
5. Produce a one-page decision brief with a recommended owner, measure, review date, and stop condition.

Definitions: [insert]
Summary table: [insert]
De-identified contact-note theme counts: [insert]
Recent process changes: [insert]

The wording matters. “Possible operational explanations” stops the output from turning an association into a claim about why a particular group did not attend. Treat contact-note themes as signals for service design, not as facts about all patients.

5. Turn findings into a decision brief

Edit the output into a brief that the service lead can approve or reject. Keep it to one page. Each proposed change needs a measurable test, not a general intention such as “improve communications”.

Use these headings:

1. Decision requested: approve one or two tests.
2. Observed pattern: the rate, count, comparison group, and review period.
3. Interpretation limits: what the data cannot show.
4. Proposed test: for example, send an existing reminder at a different approved time for one appointment type.
5. Owner and start point: who changes the process and when.
6. Success measure: no-show rate, late-cancellation rate, released-slot reuse, or contact completion rate.
7. Stop condition: when to pause because results worsen, workload is excessive, or data quality fails.

6. Check where the output may be wrong

Review every claim against the table. The model can write a plausible explanation that the figures do not support. It can also miss a process change that affected the data, such as a reminder-system outage or a change in outcome coding.

Use this check before sending the brief:

If you see this	Treat it as	What to do
A rate without a count	Incomplete evidence	Add the denominator or remove the claim
“Caused by” or “patients prefer”	Unsupported causal language	Replace with “was associated with” or remove it
A high rate in a small group	A signal, not a conclusion	Monitor for another period
Missing reminder status	A data-quality issue	Audit the reminder record before changing policy
A proposed change with no measure	An untestable action	Add a baseline, target direction, and review date

Ask a scheduling lead or booking-team representative to check whether the suggested action matches the actual workflow. Ask the service lead to confirm that the proposed measure does not create a harmful incentive, such as recording late cancellations differently to improve a reported no-show rate.

When the review does not work

If the findings are generic, narrow the dataset to one appointment type and add the process facts that were missing, such as reminder coverage or booking route. If the numbers do not reconcile, stop interpretation and repair the outcome definitions and extract first. If the proposed tests cannot be delivered within current staffing or system constraints, record that clearly and ask the service lead to choose between changing the constraint or selecting a smaller test.

Run the same review structure in the next period. Compare the test group with its own baseline, document what changed, and retire hypotheses that are not supported.