No-Show Prediction: Spot High-Risk Visits Before They Happen

Not every no-show is equally predictable, and most practices still treat every appointment on the schedule the same way: one reminder text, one no-show rate, one shrug when a chair sits empty. The research says otherwise. Across 105 published studies covering every medical specialty, no-show rates average 23% but range from 4% to 79%, and the same handful of factors keeps showing up as the reason some visits are far riskier than others.

This is not another reminder-tactics post. Cadence, two-way texting, and deposits already have a home in the reminder playbook this builds on. This post is about the more analytical layer underneath: which specific appointments on tomorrow's schedule actually carry the risk, and what a practice can deliberately do about the highest-risk ones, including the calculated bet of overbooking a slot on purpose.

The honest version includes the downside. A machine-learning model built on 4.3 million ambulatory visits reached a real, measurable accuracy edge over guesswork, and a decades-old clinic simulation shows exactly what overbooking costs when the prediction turns out wrong. Predicting risk is the easy part. Responding to it without creating a new problem is the actual skill, and that is what the rest of this guide walks through.

What actually predicts a patient no-show?

A patient no-show is rarely random. The research consistently points to the same handful of signals: a long lead time between booking and the visit, a prior no-show or late cancellation on file, being a new patient, living farther from the clinic, and booking a Monday morning or a heavily prepped visit type.

Published research, not a guess: the same two factors keep topping every study that tests them.

The most rigorous evidence comes from a 2018 systematic review that screened 727 studies and analyzed 105 of them across every medical specialty. Lead time showed up as a significant predictor in 84% of the studies that tested it, and prior no-show history in 88%, the two strongest and most consistent signals in the entire body of research. Distance to the clinic was significant in 61% of studies and day of the week in 48%, both real but less consistent than the top two.

This guide assumes the basics are already running: a normal reminder cadence in place before layering prediction on top, since prediction only pays off once forgetting is already handled. None of this means a young, new, far-away patient with a prior no-show will definitely miss the visit. It means the odds shift enough that the visit deserves a different response than the routine recall booked by a patient who has shown up on time for five years straight.

How accurate is a predictive no-show model, really?

A predictive no-show model does not forecast the future, it estimates odds from patterns in scheduling data. The best published models beat a coin flip by a wide margin, but they stay probabilistic. Treat a high-risk flag as a reason to intervene differently, never as a certainty you diagnose or announce to the patient.

The clearest real-world test comes from a 2023 study in the Journal of General Internal Medicine, built on 4.3 million ambulatory appointments from 1.2 million patients across two years. An XGBoost model reached an AUC of 0.768, noticeably better than a logistic regression model at 0.714 and far better than the health system's existing rule-based tool, which scored only 0.541. The overall non-arrival rate in that dataset was 17.7%, close to ranges reported elsewhere.

An AUC of 0.768 means the model correctly ranks a random no-show above a random attendee about 77% of the time, not that it is right 77% of the time on any single visit. That distinction sets what a practice should actually do with the output: adjust the response for a cluster of high-risk visits, not single out one patient with a certainty the data does not support.

Turn the predictors into a per-appointment risk flag

Building a risk flag does not require a data science team. Pull the five factors the research keeps confirming, lead time, prior no-show history, new-patient status, appointment type, and day or time, into a simple checklist or a scored spreadsheet column, then sort tomorrow's schedule by the count before deciding who gets extra outreach.

Six fields most practice software already stores. The flag is arithmetic, not new technology.

Most practice-management and scheduling systems already store the raw fields: booking date, visit date, patient history, and appointment type. The work is deciding which combinations count as high risk for your own practice, since a 5-day lead time might be unremarkable for a walk-in-heavy urgent care and a real outlier for a med spa that books 6 weeks out. Calibrate the cutoffs against your own last 90 days before trusting them.

A prior no-show is the single easiest flag to automate, since it is a fact already sitting in the chart, not a prediction. Practices that tag any patient with one no-show in the last 12 months and route them to a more persistent confirmation sequence often find the flag does most of the work by itself.

Which slots cluster as the highest risk

Risk is not evenly spread across the week. Long lead times compound with Monday morning slots and with visits booked after a scheduling gap, and the same handful of predictors keep clustering together in the published research: prior no-show history and lead time are confirmed far more often than day of week or appointment time alone.

Prior no-show and lead time top every list. Day of week matters least on its own.

An August 2025 MGMA member poll named the same short list practice managers see on the ground: prior no-show history, long lead time, Monday morning appointments, distance to the clinic, and visits that need heavy prep. Weather adds a smaller but real signal too. A 2024 study of over 1 million primary care visits in Philadelphia found missed appointments rose 0.72% for every 1°F below 39°F, and 0.64% for every 1°F above 89°F.

Stack two or three of these factors on the same slot, a new patient booked 5 weeks out for a Monday 8 a.m. visit with a prior no-show on file, and you have found the appointment that most deserves a different plan than the rest of the day.

What smart, targeted overbooking actually looks like

Overbooking means deliberately scheduling more patients into a session than its normal capacity, sized to the predicted no-show rate rather than guessed. At a 30% no-show rate, a full-strength overbook can recover close to a 43% increase in effective capacity, but only if it targets the slots actually carrying that risk.

The math behind this is old and well studied in operations research: if a slot shows up only 70% of the time, booking at 1 divided by 0.70 fills the gap the no-shows leave, in this example about a 43% capacity gain. One frequently cited real-world case, a community mental health clinic in Denver, used exactly this kind of overbooking to serve 157 additional patients a year, roughly one extra person seen every working day, without adding a single provider.

The targeted version applies that math only to the slots your risk flag actually surfaces: a high-demand, hard-to-refill block with two or more risk factors stacked on it. An AI receptionist can run the lighter-touch response automatically once a visit is flagged, sending the extra confirmation message on its own, but deciding who gets overbooked stays a scheduling policy call, not something to hand off blind.

No single response fits every risk profile. The table exists because the honest answer is "it depends on the slot," not a blanket rule to overbook everything that looks risky.

The real trade-off when an overbooked prediction is wrong

Overbooking is a calculated bet, not a free lunch. A published clinic simulation found that adding one overbooked appointment at a 90% show rate avoided patient wait time entirely but added about 18 minutes of average provider overtime, and the same research warned that heavy block-booking at low show rates creates real congestion once more patients arrive than predicted.

That bet does not always pay off.

The same research that produced the 43% capacity figure also modeled the downside directly: compressing appointment slots instead of adding a whole extra one trims the overtime risk, down to roughly 12 minutes of average maximum patient wait in the models tested, at the cost of a smaller capacity gain. There is no overbooking setting that erases the trade-off, only ones that manage it more or less deliberately, which is exactly why it belongs on the highest-risk, hardest-to-refill slots and nowhere else.

Put the plan to work without a data science team

None of this requires new software on day one. Start with the fields already on hand, lead time, prior no-show flag, and appointment type, sort tomorrow's high-risk slots, and route each one down a simple decision path: standard reminders, a targeted confirmation touch, or, for the hardest-to-refill blocks, a deliberate small overbook.

Sort by risk and slot demand. Every high-risk appointment gets a deliberate backup plan.

Walk the path once on paper before automating anything. A visit with multiple risk factors on a high-demand block earns the small overbook or a confirmed waitlist hold. The same risk profile on an easy-to-refill slot gets a targeted extra confirmation touch instead, since there is no capacity benefit to overbooking a slot you could fill same-day anyway. Anything still unconfirmed close to the date goes to the waitlist as the backup, not to chance.

This same predict-and-respond logic matters even more for recurring visits, since a missed chronic-care or annual wellness follow-up compounds quietly instead of showing up as one dramatic miss; see keeping recurring visits actually kept for that version of the playbook. If you would rather size your own no-show cost before building any of this, the free Growth Leak Audit puts a number on it from your own schedule, not a vendor average.