AI operating room scheduling uses machine learning to predict how long each surgery will take, find unused block time and suggest how to fill it. Studies show models usually predict case duration better than surgeons’ estimates or historical averages. The gains are real but limited by politics: who owns operating room time matters more than how accurately it is forecast.
A minute of operating room time is not cheap. A study of California hospitals, published in JAMA Surgery in 2018 by Christopher Childers and Melinda Maggard-Gibbons, put the mean cost at about $36 to $37 per minute, using 2014 financial data. That is more than $2,000 an hour, whether a patient is on the table or the room is empty while a team waits for a late start.
Operating rooms are therefore one of the most valuable assets a hospital owns. They generate a large share of revenue, determine surgical waiting lists and set the rhythm for recovery units, intensive care and wards downstream.
They are also run on a calendar that almost everyone in the hospital agrees is inefficient and almost nobody is willing to change. That is the tension any AI scheduling tool walks into. The technology is good at predicting how long operations take. The operating room schedule is not mainly a prediction problem. It is a negotiated settlement among surgeons, departments and administrators about who gets the hospital’s most valuable time, and it is defended accordingly.
How is operating room time usually scheduled?
Most hospitals, particularly in the United States, allocate operating room time in blocks. A surgeon, group or service line owns a room on a given day, say every Tuesday from 7:30 a.m. to 3:30 p.m., and fills it with their cases. Time that is not booked by a certain deadline may be released for others to use. Some rooms are kept open for add-on and emergency cases.
Blocks exist for good reasons. Surgeons need predictable time to plan clinics and cases, and hospitals need committed volume to staff rooms. But they also make the calendar sticky. A block, once granted, tends to be treated as property. Surgeons hold time they may not fill in case they need it, release it late, or keep it because giving it back might mean losing it for good. Unused block time is expensive, but reallocating it is a political decision that someone has to own.
Within each block, cases are booked using an estimated duration, typically the surgeon’s own estimate or a historical average for that procedure. Those estimates are often wrong in predictable ways. Long cases run over; short ones leave gaps. The last case of the day either finishes after staff should have gone home or is cancelled.
What can AI do for operating room scheduling?
The most mature application is case duration prediction. Instead of using a surgeon’s estimate or a simple average, a model uses dozens of variables: the procedure, the surgeon, the patient’s age, weight and conditions, the anaesthesia type, the team and even the time of day. A 2023 scoping review in BJS Open looked at 14 studies comparing machine learning predictions with standard approaches. Thirteen reported on accuracy, and seven found a statistically significant improvement. The authors concluded that the models appear more accurate but flagged significant methodological limitations in most studies, and said more work is needed on how to implement them. Evidence of efficiency savings came from just one randomised trial.
The second application is finding and releasing unused time. Models predict, days or weeks in advance, which blocks are unlikely to be filled, and prompt owners to release them earlier so other surgeons can book. Commercial tools in this space typically present it as a marketplace: a surgeon can see open time across the hospital on a phone app and request it.
The third is day-of-surgery sequencing and downstream coordination: ordering cases to reduce overruns, predicting when a patient will need a recovery bed or an intensive care bed, and flagging when the day’s plan will collide with ward capacity. This connects the operating room to the rest of the hospital’s flow, the problem we explored in Predicting Hospital Capacity Is the Easy Part. Acting on the Forecast Is Where Hospitals Fail.
Why don’t better predictions fix the schedule?
Because the obstacle is not usually information. It is ownership.
Consider what a good duration model actually tells a hospital. It will say, correctly, that some surgeons’ cases routinely run longer than booked and others’ routinely run shorter. It will show that certain blocks are consistently underused. It will predict that a given Thursday afternoon will overrun and a given Monday will leave a room idle by 1 p.m.
Every one of those findings is an implicit judgment about someone’s time, performance or entitlement. The surgeon whose estimates are always short does not want an algorithm to book fewer cases in their day. The department with an underused block does not want to lose it. The chief of surgery does not want a fight with a senior colleague who generates a large share of revenue. So the prediction is shown on a dashboard, discussed at a perioperative committee, and gently ignored.
AI does change one thing: it makes the inefficiency visible and hard to deny. Before, a surgeon could dispute a manager’s impression. It is harder to dispute two years of data on every case. That visibility is valuable, but only if the hospital has a governance process that is willing to act on it.
A worked example: the Tuesday block
Take a hypothetical orthopaedic group that holds four blocks a week at a mid-sized hospital. Its surgeons estimate knee replacements at 90 minutes each and book four per block. A duration model, trained on the hospital’s records, predicts they take a median of 105 minutes for two of the surgeons and 80 minutes for the other two.
For the slower surgeons, four cases per block means routinely running past the scheduled end, with staff overtime and occasional cancellations of the last case. For the faster surgeons, it means finishing early and leaving an expensive room idle for part of the afternoon.
The model’s recommendation is simple: book three cases for the first pair, and allow the second pair to add a shorter fifth case or release time. At $36 a minute, even 30 minutes of reclaimed time per block across 200 blocks a year is worth roughly $216,000 in operating room cost, before any added revenue from extra cases.
Now the politics. The first pair of surgeons resists booking fewer cases, because their income and their waiting lists depend on volume. The second pair does not want to release time they might need for complex cases. The service line leader worries that sharing the data will damage relationships. After six months, the hospital has adopted the model’s durations for scheduling nurses and anaesthetists but not for booking cases. It has bought accuracy and put it in the wrong place.
The numbers are illustrative. The pattern is one that perioperative managers describe constantly.
Is the politics argument too cynical?
The strongest counterargument is that surgeons are not the obstacle but the most important users. Block time exists because surgery is not a commodity; a complex spinal case needs the right surgeon, team, equipment and implants in the same room, and predictable time is what makes that possible. From this view, the algorithm’s efficiency frame misses what matters clinically. A room left idle for 40 minutes is less harmful than a surgeon rushing to fit an extra case, and a model that treats every minute as fungible will push in unsafe directions.
There is truth in that. Surgical productivity measures can create pressure that undermines safety and teamwork, and operating rooms depend on a culture where people speak up when something is wrong, a point we made in Psychological Safety in Surgery Is Not About Being Nice. An AI tool that surgeons experience as surveillance will damage that culture.
But the counterargument supports a better design, not no design. The purpose of duration prediction is not to squeeze more cases into a day; it is to book days that actually fit, so the last case is not cancelled after the patient has fasted since midnight and the team is not asked to stay two hours late. Accurate scheduling is a safety measure as well as an efficiency one. The goal is to align surgeons’ interests with the model’s use, not to use the model against surgeons.
What this means for hospitals and surgical leaders
For hospital executives, the prerequisite is a governance process for operating room time that exists before the software: clear rules for block allocation, release deadlines and reallocation based on use, decided by a perioperative committee with real authority and surgeon representation. Without that, AI will generate reports that nobody acts on.
Start with uses that help surgeons directly. Accurate case durations reduce late finishes and last-minute cancellations, which surgeons dislike as much as anyone. Tools that make open time easy to find and request help surgeons who want more time. Those wins build trust for the harder conversations about underused blocks.
For surgical leaders, share the data transparently and early, with individual surgeons seeing their own predicted and actual durations before anyone else uses them. People accept an algorithm more readily when they have been able to check it against their own experience.
For perioperative nurses and anaesthetists, accurate schedules are a working conditions issue. Predictable finishing times and fewer overruns matter for staff retention, and staff should be part of evaluating whether the tool delivers them.
For patients, the most important outcome is simple: fewer cancelled operations on the day of surgery, and shorter waits. Hospitals should report both before and after deploying AI scheduling, because those are the numbers that matter to the people waiting.
Frequently asked questions
How does AI improve operating room scheduling?
AI mainly improves scheduling by predicting how long each case will take more accurately than surgeon estimates or historical averages, and by forecasting which block time will go unused so it can be released earlier. It can also sequence cases and predict demand for recovery and intensive care beds. The gains depend on whether the hospital changes bookings and block allocations in response.
How accurate is machine learning at predicting surgery duration?
Studies generally find machine learning models more accurate than standard methods. A 2023 scoping review in BJS Open found that seven of 13 studies comparing accuracy showed a statistically significant improvement over industry standards. The review also noted significant methodological weaknesses in most studies, and accuracy varies by specialty and by the quality of a hospital’s own data.
How much does operating room time cost?
A 2018 study in JAMA Surgery, using 2014 financial data from California hospitals, estimated the mean cost of operating room time at roughly $36 to $37 per minute, in both inpatient and outpatient settings. Actual costs vary by hospital, region and procedure type, and the figure covers the hospital’s costs rather than surgeon or anaesthesia fees.
What is block scheduling in the operating room?
Block scheduling allocates fixed operating room time, such as a room on a set weekday, to a surgeon, group or specialty, who then fills it with cases. Unused time is usually released by a deadline for others to book. Blocks give surgeons predictable time but can leave rooms underused when owners hold time they do not fill.
Can AI reduce surgery cancellations?
It can help. More accurate duration predictions reduce overbooked days, which are a common reason for cancelling the last case, and earlier visibility of ward and intensive care capacity can prevent cancellations caused by a lack of beds. Cancellations also have clinical and staffing causes that AI cannot fix, so hospitals should track the reasons as well as the rate.
The operating room calendar is the hospital’s most valuable document and its most defended one. AI can tell you exactly how it is wasted. Only leadership can decide to change it.
Related on the GHC blog: Primary Care Is the Cheapest Hospital Bed, But Only If You Close One and AI Diagnostics: The Evidence Is Strongest Where the Question Is Narrowest.
Has better data ever changed how your hospital allocates operating room time? Tell us what it took.
