Ask any dental receptionist who’s going to cancel this week and they’ll probably name three people without even looking at the schedule. They’re usually right, too. The problem is that a gut feeling at 7:45 in the morning, with two lines ringing and a patient at the counter asking about insurance, rarely turns into action. Nobody has time to double-confirm a hunch. This is the first post in our Clinic Scheduling & Recovery series. We’re building a scheduling demo for small US dental clinics, and before we add a single “smart” feature, we wanted to answer a plain question: can software put a number on that gut feeling? How common are dental no-shows? More common than most owners like to admit. Adit, a dental practice software company, puts the average US dental no-show rate at around 15%, with some practices seeing rates as high as 30%. That’s a vendor number, so it’s worth taking with a pinch of salt, but the academic research doesn’t paint a prettier picture. A study of pediatric dental visits at an academic clinic in Buffalo, New York found that adolescents aged 12 to 17 had the highest no-show rate of any age group, at 24%. Think about that for a second. Nearly one in four teen appointments, gone. So the problem is real. The question is whether it’s predictable. What actually predicts a no-show? This is where it gets interesting. A 2022 paper in PeerJ Computer Science built machine learning models specifically for dental no-shows, and its best model reached an AUC of 0.718. In plain English, that’s decent but not magic. It beats a coin flip by a clear margin, but it’s nowhere near good enough to cancel anyone’s appointment on the model’s say-so. The same paper made a point we keep coming back to. Dental appointments average about 48.7 minutes, compared with 17.4 minutes for primary care, which is exactly why a dental no-show hurts more and why blind overbooking doesn’t really work. You can’t squeeze a spare crown prep into a gap the way a GP squeezes in a quick check-up. The researchers also gave a patient’s recent attendance more weight than older visits, which matches what every front desk already knows: the best predictor of whether someone shows up is whether they showed up last time. What are we trying to find out? Our research question is easy to say and harder to answer. Which booking and patient signals best predict a late cancellation or no-show at a small US dental clinic, and how early can we flag it with useful accuracy? In practice, that means finding the five signals most closely linked to cancellations and no-shows in our demo data, then turning them into a simple low, medium or high risk score that a receptionist can read without any training. We also want to know how many days ahead that score stays reliable. And just as importantly, we want to be sure it isn’t unfair to any group of patients, which we’ll come back to below. How are we testing it? Nothing exotic. We’ll start with six to twelve months of appointments from a pilot clinic, and until we have one, we’re seeding the demo with realistic synthetic data so the pipeline is ready on day one. Every appointment gets labelled with one of four outcomes: attended, cancelled early with more than 48 hours’ notice, cancelled late, or no-show. From there we’ll start simple, with logistic regression, because it’s explainable. If an office manager asks why Mrs. Patel has been flagged, we want a better answer than “the algorithm said so.” Only after that will we try a gradient-boosted model and see whether the extra accuracy is worth the lost clarity. We’ll train on older data and test only on the most recent two months, because otherwise we’d just be memorising history instead of predicting the future. At the end, we’ll report one honest number: of the appointments we flag as high risk, how many actually fall through? That matters more than any fancy metric. If we flag ten slots and only two cancel, the front desk will stop trusting the flags by Friday. What data does this need? For every appointment, the demo logs when it was booked and when it’s scheduled for, which gives us lead time, one of the usual suspects. It records the appointment type and length, the provider, and the day and hour, because some slots are simply riskier than others. It keeps a running count of the patient’s past shows, cancellations and no-shows, which is probably the single strongest signal we have, along with whether they’re new or returning, since new patients have no history to go on. Insurance is stored only as a broad category, never as policy details. The demo also notes whether a reminder was sent and whether the patient confirmed it, which tells us if the nudge worked. And finally, it records the outcome of the appointment, which is the thing we’re actually trying to predict. Is a risk score fair to every patient? This is the part most no-show tools skip, and it’s the part we care about most. If a model learns that Medicaid patients cancel more often, and the clinic responds by quietly giving them worse slots, the software has just made care harder to reach for the people who already struggle most to get it. That’s not a feature. It’s a lawsuit waiting to happen, and frankly, it’s just wrong. So we’ll check the score’s accuracy separately across patient groups, and the score will only ever trigger extra help, like a second reminder, a phone call or a waitlist backup. Never a penalty. Why this matters for a clinic Imagine the office manager opening the dashboard on Monday and seeing that three Thursday slots are likely to fall through. She calls those patients on Tuesday, confirms two of them, and lines up a waitlist patient for the third. That’s not AI hype. It’s just giving a good receptionist her gut feeling back, with a few extra days of warning.