Can Google searches predict how busy your clinic will be?

Walk-in clinics live with a hard staffing problem. Patients do not book ahead, labor is the biggest line on the budget, and the gap between a quiet Tuesday and a packed Monday shows up as either idle providers or a full waiting room. Most clinics staff to a fixed template and adjust by feel.

Our Managing Partner, Dr. Ryan Burge, spent his Doctor of Engineering research at The George Washington University on a narrower version of that problem: can a clinic forecast its daily walk-in volume well enough to staff to it, and which signals actually help?

What the research tested

The study used daily patient visit counts from a large urgent care network with clinics across many U.S. states. It screened a long list of possible predictors: day of the week, holidays, population density and income around each clinic, weather, the volume and rating of online reviews, and Google search interest for a set of health-related terms. The signals that held up were then modeled with traditional forecasting methods (multiple linear regression and seasonal ARIMA) and with machine learning methods, including decision trees, random forests and neural networks.

Several obvious suspects did not earn their place. Weather showed weak correlation with daily visits and was dropped. Online review volume did track with how busy a clinic was overall, but it could not be lined up day by day, so it could not drive a daily forecast.

What the search data added

The calendar on its own was a modest predictor. A regression using only weekday and holiday information explained about 34 percent of the variation in daily visits. Adding Google search interest for a handful of terms raised that to about 78 percent. One term, “doctors near me,” explained more than half of the variation by itself.

The pattern held for every method tested. A random forest model went from explaining about 33 percent of the variation in held-out test data to about 87 percent once search terms were included. The best performer, a seasonal ARIMA model with weekly seasonality, reached 92.7 percent, with an average forecast error under 7 percent. That compares well with published work on emergency department forecasting, most of which relies on calendar and weather inputs alone.

The calendar still mattered in a familiar way. Mondays were consistently the busiest day and Sundays the slowest, which matches what most clinic managers already know from experience.

Why this is practical for a small clinic

The useful part is not the modeling technique. It is that the strongest new signal is free. Google Trends publishes search interest by region and by week, so a clinic does not need a data vendor or a new system to use it. Combined with the clinic’s own visit history, it supports a short-range forecast for the coming days and weeks, and provider and front-desk schedules can be set to match.

The limits are worth stating plainly. The research covered one urgent care organization and walk-in demand, not scheduled appointments. Search interest is a short-range signal: it helps with the next few days and weeks, not next year’s budget. And a forecast is only as good as the visit history behind it, which is often the first hurdle for a small clinic whose records live on paper or in a system that is hard to export.

The broader lesson shows up in every industry we work in. The data that predicts demand is often sitting in public view. The organizations that use it are not the ones with the biggest analytics teams; they are the ones that asked a specific question and tested the answer.

Source: R. C. Burge, Forecasting Urgent Care Patient Volume using Traditional Techniques and Machine Learning to Improve Resource Allocation, Doctor of Engineering praxis, The George Washington University, School of Engineering and Applied Science, 2023.