The rate is a symptom. The lead time is usually the disease.
Pull your last two years of appointments, compute the number of days between when each was booked and when it was scheduled, and plot the no-show rate against that. Almost every clinic that does this finds the same shape: appointments booked for this week are kept at a high rate, and the rate falls steadily as the booking horizon lengthens, often severely past a month. The patient who books six weeks out is a different patient by the time the day arrives. Their symptom resolved, or worsened and got handled elsewhere, or their shift changed, or they simply forgot in a way no reminder fixes. Before spending a dollar on prediction, look at that curve, because if your third-next-available appointment is thirty-eight days out, you already know most of what is causing your no-show rate and no model will tell you anything more useful.
The rates themselves vary enormously by setting, and comparing yourself to a headline figure is a waste of time. Primary care commonly sits somewhere in the ten to twenty percent range. Community and safety-net clinics often run twenty to thirty-five. Behavioural health is frequently the highest in any organisation. Procedural and surgical specialties are usually lowest, partly because the appointment matters more to the patient and partly because those clinics already work the phones hard. Compare each of your own clinics to itself over time and to its own booking-horizon curve. That is the only comparison that will tell you something you can act on.
You are probably here because
- Providers sit idle at eleven while the phones say the next opening is five weeks away
- A reminder service was turned on, the no-show rate barely moved, and the contract renews soon
- Somebody has proposed overbooking and the clinicians are against it
- Two clinics in the same group report no-show rates that differ by fifteen points and nobody trusts either number
The last one is usually a definition problem, covered in the data section. The first is the backlog. The second is reminder design. The third has real arithmetic behind it and a line worth not crossing.
What a missed appointment actually costs
Two numbers get quoted and both are wrong. The first is the full visit charge, which overstates it, because the cost of an empty twenty-minute slot is the marginal contribution you lost, not the gross charge. The second is nothing at all, on the grounds that staff were being paid anyway, which understates it badly, because the slot was capacity you cannot store.
Build it from your own numbers. Take the average net collection for that visit type, subtract the variable cost of delivering it, and you have the contribution lost when the slot goes empty and cannot be refilled. Then add the part that is easy to forget: the patient who could not get in. In a clinic with a five-week backlog, an empty slot on Tuesday is not just lost revenue, it is care somebody needed and did not receive, and it is a person who will call again or go somewhere else.
Do this per visit type, because the spread is large. A short established-patient follow-up and a new-patient consultation with an hour blocked out are not the same loss, and a procedure room going empty is in a different category again. Once those numbers exist, every later decision, including whether to overbook and where, becomes arithmetic instead of argument.
Convert no-shows into cancellations
This is the frame that makes the rest of it obvious. A patient who is not coming is not the problem. A patient who is not coming and does not tell you until the appointment time is the problem, because the slot is unrecoverable by then.
So design every touchpoint to make cancelling easy and fast rather than to pressure attendance. That means a reminder with a one-tap cancel, not a reminder that says please call the office during business hours. It means no shaming language, no fee threats in the message, and a genuine offer to rebook in the same interaction. Clinics resist this because it feels like inviting cancellations. It is, and that is the point: a cancellation forty-eight hours out is worth something and a no-show is worth nothing.
Watch the pair of metrics together after you make this change. The no-show rate should fall and the advance-cancellation rate should rise. If total attendance holds and the mix shifts from no-show to cancellation, the change worked, even though one of the two numbers looks worse.
The waitlist is the highest-return thing you can build, and it needs no model
The value of a cancellation is entirely in whether you fill the slot. Most clinics fill some fraction of them through a staff member with a paper list and a spare twenty minutes, which means the slots that open on a busy morning stay empty.
What works is a short, honest waitlist and an automated offer. Patients who want an earlier appointment opt in and say which days and times they can actually manage. When a slot opens, the system texts a small group in priority order with a specific offer, first reply takes it, and the rest get a note that it went. Offer to a handful at a time rather than blasting the whole list, or you will create a race that annoys everyone who loses.
Two design details decide whether it works. The offer must be specific: a day, a time, a provider, and a link that books it. And the window must be honest about travel: an offer for a slot ninety minutes from now is unusable for most people, and a clinic that sends those trains its patients to ignore the messages. Look at your own cancellation timing distribution and set the horizon from it.
Backfill also changes what a reminder is for. If you can reliably fill a slot released two days out, then the goal of the reminder is not attendance at all, it is an early answer either way.
Reminders: send fewer, and make them answerable
Reminder programmes reduce missed appointments, and the effect is real but modest, usually a few percentage points rather than a transformation. The marginal third reminder does very little. What varies enormously is design, and that is where the remaining value sits.
Text works better than voice for most populations and both work better than email. Timing that clinics find effective is one message at booking confirmation, one around a week out where the horizon allows the slot to be refilled, and one the day before. More than that reads as nagging and people stop looking.
The response is the deliverable, not the send. Measure the fraction of patients who actively confirm or cancel, not the fraction who received a message. A patient who confirms is much more likely to attend, and an unanswered reminder is itself a useful risk signal that belongs in tomorrow's call list.
Two constraints worth naming. Privacy rules limit what a text may contain, so the message carries a time and a place and never clinical detail, and your reminder vendor needs a proper data-handling agreement in place before a single phone number leaves your system. And phone numbers rot: a meaningful share of the numbers in any practice management system are wrong or disconnected, which is the quiet reason many reminder programmes underperform their promise. Report delivery failures back to the front desk as a work queue and the programme improves without any change to the messaging.
| Lever | What it changes | Effort | Honest expectation |
|---|---|---|---|
| Cut booking lead time backlog reduction, same-day capacity | The largest driver of the rate | High — capacity and template redesign | The biggest available effect, and the hardest to execute |
| Automated waitlist backfill | Recovers revenue from cancellations | Low to moderate | Often the fastest payback in the list |
| Reminders with one-tap cancel | Shifts no-shows into cancellations | Low | A few points, plus recoverable slots |
| Ranked call list for tomorrow | Focuses limited staff time | Moderate | Works only if someone actually calls |
| Transport and support outreach | Removes a real barrier for some patients | Moderate | Large effect on a small, identifiable group |
| Overbooking by predicted risk | Fills the seat, adds crowding risk | Moderate | Use narrowly, with rules; see below |
What prediction is actually good for
A no-show model is not hard to build and it will work. Prior attendance history dominates, booking lead time is next, and after that come appointment time of day, day of week, distance from the clinic, whether the patient is new, whether they confirmed the reminder, and how many previous appointments they have rescheduled. A model on those features will separate risk clearly enough to be useful.
The mistake is turning that into a dashboard. Nobody in a clinic has time to look at a risk dashboard. The deliverable is a list: the twenty-five appointments in tomorrow's schedule most likely to be missed, ranked, with a phone number, in whatever tool the front desk already has open. A scheduler with ninety minutes works down the list, and each call has one goal, which is a clear yes or a clear no with a rebooking offer attached.
Measure it as an operational programme, not as a model. Calls attempted, contacts made, appointments confirmed, appointments released early, slots refilled. The model's accuracy is a means; the number of recovered slots is the result. We have seen good models produce nothing because the list went to a report nobody opened, and simple risk rules produce real money because a named person worked them every afternoon at three.
Predictive strength of common features — our ranking
Our ranking from the scheduling work we do, not a measurement of your panel. Order matters more than the values, and the top two are usually most of the separation.
The overbooking arithmetic, and the line we do not cross
Overbooking is a real lever and it is easy to get wrong in both directions. The comparison is between the contribution lost when a slot is empty and the cost incurred when both patients arrive, which is a longer wait, a rushed visit, staff overtime, and a clinician who did not agree to it. Because the second cost falls on people rather than on a line item, it gets underweighted.
Where it works: short established-patient visit types, in clinics with a genuine ability to absorb an extra patient, applied to slots rather than to named individuals, and capped so no session can be overbooked more than a small number of times. Where it does not work at all: procedures, anything with a room or a device constraint, and any clinic already running behind by mid-morning. A clinic that habitually finishes an hour late has no slack to sell.
The line is this. Overbooking specifically on top of patients predicted to miss concentrates the crowding cost on those same patients, and predicted risk correlates with transport difficulty, hourly work, caregiving load and distance. The patients most likely to be double-booked are the ones least able to absorb a forty-minute wait, and if they then attend, they are punished for it. We build the prediction to target support first: an earlier appointment offer, a live call, a transport arrangement, a reminder in the right language. If a clinic wants overbooking after that, do it on slot patterns and historical session-level attendance, not on an individual's risk score. It costs a little efficiency and it is the right call.
The data will be messier than the schedule looks
Every scheduling analysis runs into the same set of defects, and they are worth finding in week one rather than week six.
Status codes disagree with each other. Cancelled, cancelled by patient, cancelled by clinic, rescheduled, bumped, left without being seen, arrived late and not seen. Different sites use them differently, and the difference explains a good part of the fifteen-point gap between your two clinics. Write one definition of a no-show, in words, and get the site managers to agree to it before computing anything.
Clinic-side cancellations contaminate the label. When a provider is out and forty appointments are moved, those are not patient no-shows and they must be excluded or the model learns the provider's holidays.
A rescheduled appointment can look like two appointments. Whether a moved appointment appears as one row with a changed date or as a cancellation plus a new booking changes both the rate and the lead-time calculation. Find out which your system does.
The no-show flag is set by a person at the end of the day. Sometimes it is not set at all, and the appointment simply stays scheduled forever. Count how many appointments in your history are still in a scheduled state with a date in the past; on most systems the number is not small.
Templates are owned by clinicians. Slot definitions, hold rules, and what counts as a bookable opening differ by provider and change without notice. Anything you build over the schedule must tolerate that, because it will not stop.
Getting the data out is usually the smaller half of the problem. Reading appointments from a practice management system is generally straightforward; writing back into it is where interface work, vendor terms and clinical governance land. Design so the first version reads and produces lists and messages, and only later ask for write access.
Track the third-next-available appointment weekly, per provider
Not the next opening, which is often a cancellation and flatters you, but the third one. It is the standard measure of real access and it is the number that predicts your no-show rate months ahead. If it is climbing, your no-show rate will climb behind it whatever else you do, and the answer is capacity, template design and backlog reduction rather than anything a model can offer.
Send an appointment extract and we will show you the curve.
Two years of appointments with booked date, appointment date, visit type, provider and final status — de-identified is fine — to contact@precisionfederal.com. You get back your no-show rate against booking horizon, by clinic and visit type, and what we would do first. One business day, written, no charge.
contact@precisionfederal.comWhat we find in the first week
- No agreed definition of a no-show, so two clinics in one group are not measuring the same thing
- Clinic-cancelled appointments counted as patient no-shows, inflating the rate and poisoning any model
- Reminders sent with no way to cancel except calling during office hours
- A cancellation list nobody works, so recovered slots depend on who is at the desk
- Undeliverable numbers never fed back to the people who could correct them
- Overbooking applied by individual risk score, concentrating waits on the least flexible patients
- Success measured on model accuracy rather than on slots recovered and visits delivered
A build order that respects what actually moves the number
Scheduling programme, first eight weeks
The model appears in week six, deliberately. Almost every clinic that starts with the model spends its first two months on a prediction that is then handed to nobody, and the reminder redesign and the waitlist would have paid for the whole programme by then.
When you do not need software for this
A single-site practice with two providers does not need any of this built. It needs a definition, a spreadsheet, a reminder service with a cancel link, and a named person who calls the top ten risky appointments each afternoon. That is a week of setup and it will capture most of the available benefit.
And if your third-next-available is beyond a month, the honest answer is that no-show software is not your project. The backlog is. Reducing the wait shrinks the horizon, shrinking the horizon lifts attendance, and the improvement arrives without a model at all. That work is scheduling template design, panel management and capacity, and it is a clinical operations effort rather than a software one. We will say so rather than sell you the other thing.
Build when you run several sites with different practices, when the volume makes manual calling impossible, when cancellations are frequent enough that automated backfill is worth real money, or when nobody can currently answer how many slots went empty last month. Those are the conditions where software earns its keep.
Before you go live
- One written definition of a no-show, agreed across every site
- Clinic-side cancellations excluded from the rate and from training data
- No-show rate plotted against booking horizon, per clinic and visit type
- The cost of an empty slot computed per visit type from your own collections
- Every reminder answerable in one tap, with rebooking offered in the same message
- Undeliverable contacts returned to the front desk as a work queue
- Waitlist offers specific, small-batch, and inside a realistic travel window
- Risk output delivered as a worked call list, with a named owner and a time of day
- Any overbooking set at the slot level, capped per session, and agreed with clinicians
- Reported on slots recovered and visits delivered, not on model accuracy
Bottom line
Missed appointments are mostly an access problem and an operations problem, with a small and genuinely useful modelling problem inside them. Shorten the booking horizon and the rate falls on its own. Make cancelling easy and the unrecoverable losses become recoverable ones. Build the waitlist so a released slot gets filled by a person who wanted it. Then, and only then, use prediction to decide who gets a phone call tomorrow, and use it to send help rather than to sell the same seat twice. A clinic that does the first three and skips the fourth will do better than one that does only the fourth.
Frequently asked questions
Accurate enough to be useful and not accurate enough to be certain. With prior attendance history and booking lead time you can usually separate a high-risk group whose miss rate is several times the baseline, which is all a call list needs. Anyone promising to tell you exactly who will not attend is overselling; the outcome depends on things that happen after the prediction, like a car breaking down.
They reduce missed appointments somewhat and they fall hardest on the patients least able to pay, which for many practices is the wrong trade and for some is not permitted by their payer agreements. If it is on the table, model the revenue effect against the attrition effect first, and consider whether an easier cancellation path and a waitlist would recover the same slots without the collections work and the front-desk conflict.
Typically three: a confirmation at booking, one about a week out where the horizon allows a released slot to be refilled, and one the day before. Beyond that the returns fall away quickly and patients start ignoring them. Design matters more than count, and the number that predicts attendance is how many people answered, not how many messages went out.
Yes, and the first version usually should. Reading appointments to produce call lists, waitlist offers and reminders covers most of the value, while write-back brings interface work, vendor terms and clinical governance. Add booking write-back once the programme has proven itself, when self-service rebooking becomes the obvious next step.
Reminder and waitlist changes show up within four to six weeks in slots recovered. The rate itself moves more slowly, partly because appointments booked before the change are still working through the schedule, and partly because the largest driver, your booking horizon, only shifts when capacity or templates change. Watch slots recovered first; it responds fastest and it is the number that pays for the work.