A missed appointment is the cleanest loss in a clinic’s week. The room was heated, the clinician was paid, the slot was blocked against everyone else who wanted it, and nobody arrived. Every other kind of lost revenue at least takes some cost out with it. This one takes none.
Which is why practices are being sold prediction. Score every booking for the chance it is missed, flag the risky ones, act on the flags. It is a tidy product story. It is also, on the published evidence, the last thing a clinic should buy.
01 — What happened
Non-attendance is one of the few operational problems in healthcare with a genuinely deep evidence base, and it has been measured for long enough that the useful numbers are settled.
The broadest synthesis is a systematic review by Dantas and colleagues in Health Policy, which screened 727 papers and analysed 105 of them across every specialty. It puts the average no-show rate at roughly 23%. More usefully for an operator, it identifies what actually predicts a miss: the two most commonly reported significant determinants were high lead time — the gap between booking and appointment — and a prior history of not attending.
The intervention evidence is narrower but firmer. A Cochrane review of eight randomised trials covering 6,615 participants measured what reminders do. Attendance was 67.8% with no reminder, 78.6% with a text message reminder, and 80.3% with a phone call. Text against nothing gave a risk ratio of 1.14 (95% CI 1.03 to 1.26) on moderate-quality evidence. Text against a phone call gave a risk ratio of 0.99 (95% CI 0.95 to 1.02) — no meaningful difference. Two of the studies reported the cost per attendance achieved by text was 55% and 65% lower than by phone call. The reviewers were careful: the evidence is low to moderate quality and heterogeneous.
Then the prediction layer. A rapid systematic review in JAMIA by Oikonomidi and colleagues looked specifically at interventions targeted by a risk model — eight studies, seven of them randomised. Model-targeted text reminders worked, at high certainty, with a median risk ratio of 0.91. Model-targeted phone calls (median RR 0.61) and patient navigator calls (RR 0.55) showed larger effects at moderate certainty. Model-targeted overbooking was rated very low certainty, with an uncertain effect. The authors could not pool the results, and flagged the gap that matters most here: nobody has properly compared targeting against simply doing it for everyone.
The largest trial of targeting gives a sense of scale. Ulloa-Pérez and colleagues randomised 125,076 primary care and 33,593 mental health visits, all already flagged as high risk, between one text reminder and two. The second text reduced no-shows by 7% in relative terms for primary care (RR 0.93) and 11% for mental health (RR 0.89).
02 — Why it matters
Read those studies in order and the sequence inverts the sales pitch.
The intervention with the largest measured effect is having a reminder at all: around eleven percentage points of attendance, which is roughly a third of all non-attendance removed. The intervention with the smallest measured effect is the clever one — a second message aimed by a model at the people most likely to miss, worth 7% to 11% relative on a population that has already been reminded once.
And the channel comparison kills the usual justification for targeting. Models are sold on efficiency: stop spending on everyone, spend on the ones who need it. But the Cochrane trials found the cheapest channel performed as well as the dearest. If a text costs almost nothing and works as well as a call, the thing a prediction model economises on is the input that was already nearly free.
03 — The economics
Work it through on an illustrative clinic — these figures are an example, not a finding.
Say four clinicians, twenty slots each a day, five days: 400 slots a week. At a 15% no-show rate, below the 23% review average, that is 60 empty slots a week that were fully staffed and fully costed. The clinician’s hour, the room, the reception time booking it — all spent. The revenue is gone, and so is the appointment someone else wanted.
Apply the Cochrane effect to a clinic that currently reminds nobody. Attendance moving from 67.8% to 78.6% is about a third of non-attendance removed; on 60 missed slots, call it 20 recovered a week. Nothing was added to the cost base to get them: no extra clinician, no extra room.
Now apply the targeting effect on top. Of the 40 misses that remain, suppose half sit on the model’s high-risk list, and a second message to them delivers the 7% to 11% relative reduction the trial measured. That is a fraction of one slot a week. The first reminder is worth something like twenty times the prediction layer, and far cheaper, because the model has to be built, fed with clean booking history and monitored.
Wording is the other free lever. Hallsworth and colleagues ran two randomised trials across roughly 20,000 outpatient appointments testing only the text of the reminder. The existing message produced 11.1% non-attendance. A message stating the specific cost of a missed appointment produced 8.4% (OR 0.74, 95% CI 0.61 to 0.89). In the second trial, the specific-cost message held at 8.2% while a vague reference to wasted money came in at 9.9% and an empathetic appeal at 10.7%. Nearly three percentage points, for the price of rewriting a sentence once.
The pattern is the one we described in why exceptions decide whether automation pays: the sophisticated part of the system is rarely the part carrying the return.
04 — What businesses should do
In this order, and do not skip to the end.
Measure your own rate, split by lead time and by slot type. The review names lead time as a top determinant, so a practice booking routine appointments six weeks out has a different problem from one booking three days out. The same logic applies to your retention numbers, as we argued in the retention statistics everyone quotes are from 1990 — the benchmark is interesting, yours is actionable.
Confirm every appointment actually gets a reminder, and that it fires when the booking changes. Most clinics believe they do this. Fewer survive the test of booking an appointment, moving it, and seeing what the patient receives.
Rewrite the message before you buy anything. Specific and concrete beat vague and sympathetic in both trials. Name what a missed appointment costs the practice and the waiting list, state it plainly, and make cancelling easier than not turning up.
Build the fill, not just the warning. A reminder that converts a no-show into a cancellation only pays if you can refill the slot. That means a reply to the reminder has to release the appointment, and releasing it has to message a waiting list. The research does not price this — treat it as mechanism, not evidence — but it is the difference between knowing about a gap and selling it.
Shorten lead time where it is clinically sensible, since that is where the determinant sits.
Only then consider a model — and if you have one, point it at human effort rather than more free messages. The targeted studies found calls outperforming texts by a wide margin. A flagged list is a worthwhile queue for a receptionist’s afternoon. Do not let it authorise overbooking: that was the one intervention the review rated very low certainty.
05 — Where technology fits
For most single-site practices, honestly: not AI. Not yet, and possibly not ever for this problem.
What the evidence asks for is dull and reliable. Reminders that fire automatically from the booking system, including after a change. Two-way messaging, so a reply can cancel. An integration that does not depend on someone exporting a list. A record of who missed, who rebooked and who never came back — prior non-attendance is a real predictor, and you can act on it with a note and a phone call, no model required. That is automation and integration work, the kind described in how we think about systems, and the same argument we make across healthcare and clinics: a process that runs without anyone remembering it.
A prediction model earns its place at volume. If you run tens of thousands of appointments a month across several sites, with reminders already reliable, wording already tuned and staff time to spend on a flagged list, scoring bookings starts to pay: a one-percentage-point move is worth a salary. Under a few thousand a month, the model competes for attention with two interventions that have better evidence and cost nothing.
06 — RevenueStack’s take
Our view: the interesting thing about the no-show literature is not that prediction works. It is that the two interventions with the strongest evidence — sending a reminder at all, and choosing its words carefully — are nearly free, have been documented for over a decade, and are still done badly in most practices.
That gap persists because of how things get bought. A model is a product with a vendor, a demonstration and a budget line. Rewriting a reminder is an afternoon with no invoice attached, so nobody schedules it. The result is clinics paying for the 7% while the 30% sits untouched.
Before you evaluate anything that scores a patient, book an appointment in your own system, move it, and read what the patient gets. Most of the return is in that message.
Sources
- No-shows in appointment scheduling – a systematic literature review — Dantas, Fleck, Cyrino Oliveira and Hamacher — Health Policy 122(4):412-421, 2018-04
- Mobile phone messaging reminders for attendance at healthcare appointments — Gurol-Urganci, de Jongh, Vodopivec-Jamsek, Atun and Car — Cochrane Database of Systematic Reviews, 2013-12-05
- Predictive model-based interventions to reduce outpatient no-shows: a rapid systematic review — Oikonomidi, Norman, McGarrigle, Stokes, van der Veer and Dowding — Journal of the American Medical Informatics Association 30(3):559-569, 2023-03
- Pragmatic Randomized Study of Targeted Text Message Reminders to Reduce Missed Clinic Visits — Ulloa-Pérez, Blasi, Westbrook, Lozano, Coleman and Coley — The Permanente Journal 26(1):64-72, 2022-04-05
- Stating Appointment Costs in SMS Reminders Reduces Missed Hospital Appointments: Findings from Two Randomised Controlled Trials — Hallsworth, Berry, Sanders, Sallis, King, Vlaev and Darzi — PLoS One 10(9):e0137306, 2015-09-11
