By Karen Blum
ORLANDO, Fla.—Artificial intelligence may be helpful in surgery and perioperative medicine, particularly in the setting of time constraints and uncertainty when making high-stakes decisions that are shared with patients, according to a surgeon presenting at the 2024 Perioperative Medicine Summit.
Under those conditions, surgeons tend to turn to their standard, hypothetical deductive reasoning, said Tyler Loftus, MD, PhD, an assistant professor and acute care surgeon in the Department of Surgery at the University of Florida, in Gainesville.
“In a lot of circumstances, we’re very effective in applying that,” according to Dr. Loftus. But when pressed for time and with incomplete information, surgeons fall back on cognitive shortcuts honed through practice. That can work in some cases, but without experience, those decisions can be erroneous or allow for implicit biases.
“The rationale for AI in perioperative medicine is not to replace the clinicians,” he said. “It’s only to say that we think the human with the algorithm [and its objectivity] is perhaps a bit better off than the human alone.”
Traditional decision support systems use rule-based algorithms to decide what should be done for a patient, Dr. Loftus said, citing a 1987 article (N Engl J Med 1987;316[11]:685-688). But medicine is so broad and complex that it is difficult to capture relevant information in these rules. AI models, by contrast, learn from examples, which can help because patients can present with any of thousands of diagnoses, be on any of numerous medications or undergo any of a multitude of procedures.
Dr. Loftus shared some examples of how AI can be and is being used in patient care. For example, risk calculators like the American College of Surgeons NSQIP (National Surgical Quality Improvement Program) or POTTER smartphone app allow surgeons to input preoperative information about patients to provide estimates regarding their risk for postoperative complications. Going a step further, the MySurgeryRisk calculator from the University of Florida extracts information from electronic health records automatically.
AI also could play a role in helping make postoperative triage decisions, determining whether patients should be sent to an ICU or regular ward after surgical procedures to best align the intensity of resources needed based on their status. Dr. Loftus and colleagues have a program they are looking to integrate into their electronic health record.
One strong example of how AI improved care for patients is the HYPE trial out of Amsterdam (JAMA 2020;323[11]:1052-1060), in which investigators used a machine learning–derived early-warning system to predict intraoperative hypotension in patients undergoing elective noncardiac surgery, change their behavior and decrease intraoperative hypotension, he said.
Going forward, federated learning is a methodology some researchers are moving to that can factor in data from multiple healthcare centers to devise solutions that would apply to a greater number of scenarios, without ever sharing patient identification between centers. “Not only is it collaborative learning and likely more generalizable, but it can make smarter, better models that learned from more cases and more rare scenarios,” Dr. Loftus said.
A number of ethical, legal and social issues exist in the AI sphere that stakeholders should work to overcome, Dr. Loftus said. In one study (Ann Surg 2023;277[3]:359-364), only five of 84 surgical clinical decision support systems were found to report equity analyses, like assessing how the algorithm performs on different patient populations. AI systems should be trained on balanced data sets that represent vulnerable populations from heterogeneous practice settings. Federated learning is one approach that can use information from multiple healthcare centers with different patient populations.
Additionally, omit or justify the inclusion of social determinants of health as model features, he advised. For example, if you want to predict something related to a disease much more likely to affect women, then you may want to have sex as a predictor. But if race is introduced as a social construct rather than gene-based predictor, bias could result.
Finally, ensure that predictive performance is strong across subpopulations, he said. New technologies disproportionately aid high-income countries and the Global North. To counteract that, researchers should perform needs assessments that include stakeholders and perspectives from low- and middle-income countries and the Global South, and work to build modules that will work in all scenarios.
“At the end of the day, AI is probably most useful in letting us spend less time making calculations, and more time interacting with patients and using our intuition, creativity and altruism,” Dr. Loftus said.
Dr. Loftus’ presentation was “a brilliant overview of a sometimes bewildering topic,” commented session moderator Kurt Pfeifer, MD, FACP, SFHM, DFPM, the chief of the section of perioperative and consultative medicine at Froedtert Hospital in Milwaukee, and a professor of medicine at the Medical College of Wisconsin, also in Milwaukee.
“AI will undoubtedly reshape all of healthcare, and so it’s critical perioperative medicine clinicians develop an understanding of this area,” Dr. Pfeifer said. “Dr. Loftus was able to distill a challenging topic into knowledge the entire audience could understand. His citing of previously published use of AI coupled with opportunities for perioperative use in the future were thought-provoking and hopefully made attendees energized for future AI applications.”