Predicting Surgery Durations

For years, surgeries have been scheduled according to the average amount of time a procedure takes.

But this has never been very efficient or cost-effective. In practice, surgery times vary considerably, and a good share of that variance can be explained by factors that are known in advance.

Uncertainty and Human Attributes Influence Surgery Duration

For example, attributes of the patient can make a big difference—co-morbidities, age, height, weight, and sex can all influence surgery duration. Attributes of the doctor performing the surgery can also factor in here: Some surgeons take longer; some take less time. Even the location at which the surgery is performed (either the facility or the specific operating room) can make a difference.

In addition to getting more accurate estimates of the expected surgery duration, it’s also important to estimate our uncertainty. If one surgery is known to take exactly 30 minutes, then we should book exactly 30 minutes. If another is expected to take 25 minutes but can vary wildly, then we might want to book more time to be safe.

In short, both the expected surgery duration and our uncertainty will depend on what procedure is being performed, the health characteristics of the patient, the record of the surgeon, and the site of the operation.

Including Uncertainty in Statistical Model Yields More Accurate Predictions

For example, the expected duration and the uncertainty tend to be correlated. Operations expected to take an hour might exhibit greater variance than those projected to take 10 minutes. Also, variability could be higher for complex operations than routine ones. Multiplied across many thousands of annual surgeries, the costs of overbooking or underusing hospital resources can be enormous.

A better estimate of a surgery’s duration can be made by estimating this uncertainty and factoring it into the decision-making process. In one study, we leveraged modern neural networks, a kind of statistical model that processes patterns and relationships through a cascade of layers, each of which maps the data into a space more amenable to making accurate predictions. We then used the neural network to predict how long surgeries would take and, at the same time, to estimate the variance (both conditional on the observed features).

Using technology like the approach proposed in our paper this way, health care providers might schedule surgeries more efficiently and cost-effectively while allowing an informed margin of error that is specific to each case.