UF study uses AI to predict postoperative complications, improve patient care

A successful surgical outcome is composed of two parts: What happens during the surgery, and what happens after.

Ideally, the latter ensures the technical skill exercised during the procedure has a chance to flourish. But occasionally — due to individual factors, resource availability, human error and the fast-paced nature of surgical care — postoperative complications can occur.

Now, University of Florida researchers are leveraging artificial intelligence to make accurate predictions of postoperative complications — and applying those predictions directly to patient care.

“We use the predictions of postoperative complications to determine whether a patient after surgery should be admitted to an intensive care unit with high-frequency surveillance and immediate availability of specialized personnel trained to diagnose and treat organ failure or to a hospital ward where they’ll be monitored intermittently about every four hours on average,” said Tyler Loftus, M.D., acute care surgeon and researcher.

photo of Dr. Loftus scrubbing in to the OR
Dr. Loftus’s peer-reviewed publications and research presentations at national and international meetings have evolved from translational science focusing on bone marrow failure and anemia after traumatic injury to data science focusing on machine learning to augment personalized, patient-centered decision-making. He is the Associate Director of Research for the University of Florida Intelligent Critical Care Center.

In a study published in JAMA, Loftus and his coauthors applied the prediction framework to identify patients who were being undertriaged, meaning that they had a high risk for complications, but were assigned to a low-frequency surveillance environment. Broadly, the study investigated the association of undertriage to hospital wards after surgical procedures with mortality and resource use.

“We were able to identify [the undertriaged patients] and test and confirm the hypothesis that they would have worse outcomes compared with ICU patients who had a similar risk profile,” Loftus said.

Although the link between undertriaging and patient outcomes is intuitive, there is a scarceness of concrete, data-driven evidence regarding associations between postoperative undertriage and patient outcomes. Loftus and his coauthors intend this study — published in JAMA Network Open — to be the first of several that explore ways in which electronic health record data can be used to optimize clinical judgment.

The algorithm used electronic health record data, allowing researchers to pinpoint orders being written in Epic in real time — while also providing an opportunity to modify or redirect patient postoperative triage where necessary.

A surgeon by training, Loftus, who is pursuing his Ph.D., is focused on the intersection of artificial intelligence as a broker between physicians, clinical decision-making and the ocean of health data that accompanies patients as they walk through the door.

“A computer is typically better at automatically and almost instantaneously obtaining data, running an algorithm and making a prediction that’s highly accurate in a way that is beyond human capacity,” Loftus said.

Algorithms like this one have the potential to impact care for millions, Loftus said.

            “When carefully designed and informed, real-time, machine-learning models can help us put our patients first,” Loftus said.