Every outpatient department loses slots to patients who do not turn up. The usual answer is a reminder, and our earlier write-up on predictive analytics for no-shows covers why a risk score can direct that effort. This article goes one level deeper, into the parts that decide whether a model helps or quietly harms: which features belong in it, how to test it for bias against groups of patients, and how to price a wrong prediction before you act on one.
It is written for hospital analytics leads, operations heads and the marketing or patient-experience teams who sponsor these projects. It assumes you have appointment history in a hospital information system or CRM and some ability to run a logistic regression or gradient-boosted model. Everything numeric below is synthetic, chosen to show the calculation. Nothing here is a benchmark, and nothing replaces review by your clinical and data-protection leads.
Define the prediction carefully
A no-show model answers: given what we know at booking time (and at set points before the slot), what is the probability this appointment will not be attended?
Decide four things first.
- The label.
no_show = 1 if the patient neither attended nor cancelled before a cut-off, such as two hours before the slot. Late cancellations that freed the slot are a separate class and should not be lumped in, because the intervention differs. - The prediction time. Score at booking, again at 48 hours and again at 24 hours. Features available only later, such as whether a reminder was read, must not leak into the earlier scores.
- The unit. One row per appointment, not per patient. Repeat patients appear many times, so split train and test by patient to avoid learning that Mr X always turns up.
- The use. Write down what the score will trigger. If the answer is "nothing yet", you are not ready to build it.
Features that usually carry signal
Think of features in groups, and for each ask whether it will be available when you need it and whether you are allowed to use it.
| Group | Examples | Notes |
|---|
| Booking | Lead time between booking and slot, booking channel, day and hour of slot, specialty, first visit or follow-up | Lead time is typically one of the strongest predictors; verify on your data |
| History | Previous attended and missed visits, time since last visit, past late cancellations | Strong, but see fairness below |
| Appointment type | Procedure, test with preparation, referral present | Tests needing fasting or prep carry their own risks |
| Logistics | Travel distance, clinic location, parking, public holiday or exam week | Use area or distance band, not exact address |
| Communication | Confirmation received, reminder delivered, confirmation reply | Available only after the reminder; treat as a later-stage feature |
| Payment | Prepaid, insurance pre-authorisation status | Useful, but can proxy for income |
Start simple. A regularised logistic regression with ten features is easy to explain to a head of department, easy to audit and often close to a complex model on tabular data. Move to gradient boosting only if you can show a gain on held-out data that matters operationally.
What to leave out
Exclude features that identify or proxy for protected or sensitive attributes unless you have a clear reason and sign-off: religion, caste, community, language used as a stand-in for either, disability and detailed diagnosis. Be wary of pincode and name-derived variables, which can proxy for community or income. The DPDP framework, notified in 2025 and phased in with most obligations from May 2027 as summarised here, makes purpose limitation and security safeguards your concern. Document the purpose of each field, and ask counsel what notice and consent your use needs.
Validate like a forecaster
Split by time. Train on older months and test on the latest, as in production. Report:
- Discrimination: AUC, to show the model ranks risky appointments above safe ones.
- Calibration: a reliability diagram. The scikit-learn guide to probability calibration explains how to build one and how sigmoid and isotonic calibration work. You need real probabilities, because you will multiply them by costs.
- Precision and recall at your operating threshold, not at 0.5.
- Stability: repeat by month and by clinic.
A model that is accurate on average can be wrong where it matters, which leads to fairness.
Fairness: test the errors, not the intentions
A no-show model learns from history, and history includes obstacles that fall unevenly. Patients who travel far, work shifts, rely on others for transport or have less flexible jobs may miss more appointments for reasons unrelated to commitment. If the model then gives them double-bookings or removes their favoured slot, it penalises disadvantage.
You do not need a fairness research team to start. Do this:
- Choose groups you can legitimately measure. Age band, gender where recorded, travel distance band, first visit versus repeat, insurance versus self-pay and booking channel are usually available. If you cannot lawfully analyse a sensitive attribute, use proxies carefully and say so.
- Compute error rates per group at the operating threshold. The two that matter most are the false positive rate (patients flagged as likely no-shows who would have attended) and the false negative rate.
- Compare calibration per group. If a predicted 0.4 corresponds to an observed 0.4 in one group and 0.25 in another, the model over-predicts for the second.
- Set the policy before looking. Decide what gap in false positive rate is acceptable, who decides, and what happens if it is exceeded.
Synthetic example. At the chosen threshold, the model flags 1,000 appointments from patients living within 5 km and 1,000 from patients living beyond 20 km.
| Group | Flagged | Of whom attended | False positive rate among flagged | Missed no-shows (of group's no-shows) |
|---|
| Within 5 km | 1,000 | 450 | 0.45 | 120 of 400 |
| Beyond 20 km | 1,000 | 620 | 0.62 | 60 of 300 |
Among flagged patients, those living far away were more often wrongly flagged: 620 of 1,000 actually attended. If the intervention were to double-book their slots, 620 attending patients would be competing for a place. The numbers justify a different intervention for that group, or a different threshold, or a feature change. The point is that you can see it, which you can only do by splitting the error rates.
Remove a feature only if testing shows it hurts; sometimes the correct fix is to change the intervention. Equal error rates across all groups are often mathematically incompatible with perfect calibration when base rates differ, so you will need judgement and a written rationale rather than a single magic metric. The NIST AI Risk Management Framework is a useful structure for documenting that judgement, covering how risks are mapped, measured and managed.
Price the false positive
Every flagged patient triggers an action. Each action has a cost when applied wrongly. Build a small table before choosing a threshold.
| Intervention | Cost if the patient would have attended anyway (false positive) | Benefit if the patient would have missed (true positive) |
|---|
| Extra reminder (WhatsApp or call) | Small staff or messaging cost; mild irritation | Recovers some slots |
| Personal call from coordinator | Staff minutes | Higher recovery chance, plus rebooking if needed |
| Ask for a confirmation reply | Friction; may cause some to drop out | Frees slots earlier |
| Double-book the slot | If both arrive, a waiting patient and an overloaded clinician | Fills the slot |
| Require a deposit | Deters genuine patients; access and reputational issues | Reduces no-shows but can reduce access |
Now do the arithmetic with invented figures. Suppose a call costs 40 rupees in staff time, a recovered consultation slot is worth 800 rupees in clinic revenue, and a call recovers a slot in one of every four calls to true no-shows. For a patient with predicted risk p, expected benefit is p × 0.25 × 800 = 200p rupees, and cost is 40 rupees. The call is worthwhile when 200p exceeds 40, so for p above 0.20. This yields a threshold from costs rather than from a statistic.
Do the same for double-booking, where a false positive means someone waits. Put a number on a patient waiting forty minutes for a surgeon: perhaps measured through complaints or reviews. If the cost is high, restrict double-booking to patients whose predicted risk is very high and calibrated, in specialties where short delays are tolerable, and never for time-critical care.
The structure means you may use different thresholds for different interventions: a low one for a free reminder, a higher one for a phone call, a very high one for overbooking. A single flag is rarely right.
Operations and governance
- A human decides anything adverse. Do not use scores to deny, delay or de-prioritise care. Use them to add support, such as a reminder, transport information or a call.
- Keep the score internal. Patients should not see a risk label.
- Log decisions. Record the score version, threshold, action taken and outcome for each flagged appointment.
- Run an experiment. Hold out a random group that receives standard reminders, so you can measure whether the intervention improves attendance beyond what the model implies. Without a control you cannot separate model value from seasonality.
- Monitor drift. Recheck calibration and group error rates monthly. A new clinic, a holiday week or a change in booking channel can shift behaviour.
- Retrain on a schedule, and archive each version.
If your reminders run through a CRM, Pioneer CRM is one option for keeping appointments, owners and follow-ups together; the modelling above does not depend on any tool.
Where to start
- Write the label, prediction times and intended actions.
- Pull 12 months of appointments; engineer the booking, history and logistics features; split by time and patient.
- Fit a simple, regularised model; check discrimination and calibration.
- Produce error rates and calibration by group, with a written threshold for acceptable gaps.
- Build the cost table and set thresholds per intervention.
- Pilot with a control group before any rollout.
For background, see the no-show rate entry, our no-show elimination framework and the predictive analytics service page. Your clinical and patient-experience leads should review every intervention before it goes live.
Filed underno show prediction model hospitaloutpatient no-show model featuresno show prediction fairnessappointment no-show machine learningoverbooking false positive cost

About the author
Arush Thapar
Co-Founder & CTO · Gurugram, India
Arush builds the systems the rest of the company runs on — the patient-acquisition pipelines, the AI chatbots, the analytics that tie a single click to a booked appointment. If a campaign reports a number, it's because something his team built is tracking it.
Free guide · 15 pages
The Patient Acquisition Blueprint
A 90-day plan from first search to booked appointment.