Predicting Operating Room Activity

The operating room is the most expensive room in a hospital. A single OR can generate more revenue than an entire floor of beds. So if you're running a hospital, how you use your operating rooms matters more than almost anything else on the balance sheet. The catch is that, for most of the history of American healthcare, the people in charge of scheduling those rooms were doing it with a pencil, a paper calendar, and a clipboard.

I joined a company in early 2020 that was trying to change that. Their idea was that if you could accurately predict which cases would actually show up, how long each one would take, and how the day would cascade when something ran over, you could turn that pencil-and-paper mess into something closer to OpenTable for surgery. But there was a problem. The in-house data science team had spent years trying to build the predictions and had concluded it couldn't be done. Hospital operations, they said, were stochastic. Too many variables, too much noise. Not a statistics problem anymore. A weather problem.

I had just come from three years building prediction models for industrial equipment, which everyone had also said couldn't be done, so I was skeptical of the skepticism. I spent my first few months with the actual data and the actual operations people: the schedulers, the OR coordinators, the surgeons who had been running all of this with clipboards. A lot of what they called chaos turned out to be patterns hiding in plain sight. The predictions started landing. The ORs started filling in smarter. Hospitals that installed the system reported real, measurable revenue lift.

The company ended up serving a large hospital network and was eventually acquired. I ended up reporting directly to the president of the company, which isn't something I would have predicted on day one. COVID hit three weeks after I started, so most of that work happened through a screen during lockdown. It was a weird and meaningful time.