The three o’clock list worked, more or less. Every afternoon, each of ClaraCare’s twelve clinics received tomorrow’s thirty highest-risk appointment slots, defined by three sensible rules - new patients, patients with a prior no-show, appointments booked more than a month out - and the front desk made personal calls. Over two quarters the no-show rate slid from roughly eighteen percent toward fifteen.

But the front-desk leads noticed something the dashboard did not show. Most of the thirty patients they called turned up anyway. And the no-shows that actually emptied rooms kept coming from outside the list: established patients booked a week ahead, Monday-morning slots after a long weekend, appointments with a high expected patient bill. The rules were a reasonable guess. Nobody knew how good a guess, because measuring it would have required a model to compare against and an honest test of both.

ClaraCare’s analytics function is the revenue-cycle lead and a spreadsheet. The EHR vendor offered a no-show risk score as an add-on module - trained on other organizations’ patients, with accuracy reported on the vendor’s own data. That distinction matters more than it sounds. A 2025 review of 52 studies found that the best no-show models reported AUC scores between 0.75 and 0.95. But when one outpatient radiology model was tested prospectively on real future appointments, its AUC fell from 0.93 on historical data to 0.73 - still useful, with its highest-risk group three times as likely to miss as everyone else, but a very different number. A score measured on the data it learned from is a claim. A score measured on next month is a fact.

Measured on next month

The COO wrote the request with that difference in mind:

> Predict the probability that each scheduled appointment in the next 14 days will be a no-show, using our last two years of appointments. Test it on the most recent three months, not on data it trained on, and tell me how our three current rules compare.

The Prediction Engine found about 180,000 past appointments, eighteen percent of them no-shows, and built its inputs strictly from scheduling and practice-management data: lead time, day and hour, appointment type, clinic and provider, the patient’s own attendance and cancellation history, estimated patient responsibility, and the weather. No clinical notes and no diagnoses. It split by time rather than at random - train on the first twenty-one months, test on the last three - so the result would reflect a future the model had not seen. And it scored the three existing rules the same way, as a baseline.

The engine also surfaced a choice that belongs to people, not models. Payer type was one of the strongest predictors. The COO kept it out of the production model: the scores would soon influence where extra appointments get booked, and she did not want insurance coverage steering which patients risk a longer wait. The engine re-ran without it and reported the cost - a point and a half of accuracy - so the trade-off was made knowingly and written into the specification.

On the three held-out months, the model reached an AUC of 0.78. The three rules, scored as a ranking, reached 0.61. Evaluated on its own training data, the model would have claimed 0.86; the engine shows that number only so nobody mistakes it for the real one. In practice, each clinic’s thirty highest-risk slots by model contained about twice as many actual no-shows as the thirty chosen by the rules.

Three scheduling rulesnew, prior no-show, booked >1 month 0.61

Model, on its own training datanever the number that ships 0.86

Model, held-out monthsthe number the COO sees 0.78

0.50.60.70.80.91.0

Ranking power (AUC), where 0.5 is a coin flip

Composite scenario; figures illustrative. Operational scheduling data only; no clinical records used.

The front desks switched lists the following week. Same thirty calls, same staff time, better aimed. Over the next two months the no-show rate moved from about fifteen percent to under thirteen. At ClaraCare’s volume each point is roughly fifteen thousand dollars a month of clinician time used rather than lost.

Denials, predicted before submission

The revenue-cycle lead’s denial tripwire from the Analytics Engine catches a payer changing its behaviour after eleven denials instead of a hundred. The better outcome is the denial that never happens. The Change Healthcare Denials Index found 86 percent of denials potentially avoidable, with registration and eligibility issues the largest single cause.

So she asked for a model: for every claim waiting to be submitted, predict the probability it will be denied and the most likely reason, trained on 24 months of claims and remittances. The engine validated on the most recent quarter and reported calibration by risk band, because a biller needs to trust that a claim marked forty percent risky really is. It also returned the drivers in the billing team’s own vocabulary. The imaging prior-authorisation quirk they had discovered during the payer renegotiation was the top driver for two codes.

Claims in the highest band - about eight percent of volume, holding more than half of eventual denials - now receive a three-minute check before they leave: eligibility re-verified, authorisation confirmed, modifiers reviewed. That is a small amount of work placed exactly where it pays.

0%10%20%30%40%

1.8%4.4%8.1%21%38%

Lowest41% of claimsLow26%Middle16%High9%Highest8%

Lowest41% of claims 1.8%

Low26% 4.4%

Middle16% 8.1%

High9% 21%

Highest8% 38%

0%10%20%30%40%

Predicted denial rateObserved denial rate

• Highest band: 8% of claims, 52% of eventual denialsScrubbed before submission

Composite scenario; figures illustrative. Predictions use billing, coding and eligibility data; the model suggests a likely reason code.

What the models are for, and what they are not

Both models run on operational and billing data and serve operational decisions: who gets a reminder call, which claims get a second look. Neither touches care. No-show scores are never used to refuse or delay an appointment, and the specification records that constraint alongside the features that were deliberately excluded.

The six-person revenue-cycle team gained something it never had: a way to test its own instincts. Several of those instincts held up. One - that a particular payer denied more on Fridays - did not, and the team stopped building its week around it.

The three o’clock list is now a measured list. But the empty slots it cannot prevent still exist, around thirteen in every hundred. The physicians have started asking the obvious next question: if we know roughly who won’t come, can we book someone else into that time without filling the waiting room? That is not a prediction. It is a schedule, and it has a best answer.