by Amy Keller
Every year, approximately 15 million Americans undergo some sort of surgery, and nearly one-third develop a complication. Each adverse event can add up to $11,000 in health care costs and exact a physical and emotional toll on the patient. But a promising AI tool being developed by researchers at the University of Florida could help doctors better forecast how patients will do before they even go under the knife.
The machine-learning platform, known as MySurgeryRisk, was built from the de-identified electronic health record data of more than 58,000 adult patients who underwent more than 74,000 surgical procedures at UF Health in Gainesville. UF scientists then trained the algorithms to detect patterns in that data to predict eight major complications: Sepsis (a life-threatening response to infection), wound complications, blood clots, acute kidney injury, prolonged mechanical ventilation, neurological complications such as delirium, cardiovascular complications and prolonged ICU stays.
The algorithms, it turned out, were just as good as doctors at predicting surgical outcomes — and the system has evolved into a mobile app that delivers predictions to surgeons in real time. “We are allowing them to see similar patients they’ve operated on before, (and see) how they compare to that patient and so on,” says Dr. Azra Bihorac, an ICU intensivist and lead researcher on the project. “We think this is something that can really be a great tool at the bedside and be used by physicians, and hopefully patients, to understand better what you can expect from surgery.”
Fragmented care presents an ongoing challenge to surgical risk assessment. Patients tend to see a lot of different doctors before, during and after surgery — but those physicians “don’t necessarily always talk to each other” and information is often lost in the transfer of care from one physician to another, Bihorac says.
MySurgeryRisk helps to tackle that problem by pulling all of a patient’s relevant health data into one platform that’s able to connect the dots. Those dots include 135 variables encompassing everything from the patient’s preexisting health conditions, medications, lab results to demographic data known as “social determinants of health.” The algorithm also takes into account the type of surgery being performed, as well as the surgeon’s track record and rate of complications.
While MySurgeryRisk is still in the testing stages, Bihorac and her team have high hopes for the technology. At the most basic level, better predictive information could enable patients, their families and their doctors to make more informed decisions about care. Beyond that, surgeons and other providers could one day use the AI-generated risk scores to take proactive steps to mitigate a patients’ risks in the operating room and during recovery.
Take post-operative acute kidney injury, for instance. Though often under-recognized, it’s one of the most common complications of surgery and can affect up to 50% of surgical patients, depending upon the procedure. It can also lead to long-term kidney disease as well as increase a person’s risk of developing cardiovascular disease and sudden cardiac death.
Knowing in advance that a patient has a high risk of developing acute kidney injury, the operating team could take certain steps to try to lessen any potential damage to the kidneys, Bihorac says. “You could use different types of anesthetics and avoid certain medications that are poisonous to the kidneys. You could also focus on applying different blood pressure strategies and using more individualized blood pressure medications,” she says, as well as more precisely managing intravenous fluids so as not to tax the organs. “We are really excited to push this and see in future trials how these risk assessments can be combined with standards of care and maybe we can push those boundaries and find some new ways to decrease these complications.”
The AI-generated risk scores could also help in assigning patients to the appropriate units for optimal recovery. Doctors sometimes overtriage, admitting patients to an intensive care unit for close monitoring and immediate access to specialized personnel that likely isn’t even necessary. While most do so to err on the side of caution, it’s not always a benign choice. Hospital ICUs can be havens for multi-drug-resistant bacteria, and ICU patients have a higher risk of developing hospital-acquired infections. “The idea is not only to identify who are the patients that we really need to focus on, but who are the patients we can fast-track and avoid risks — because we now know excess medical care can lead to harm too,” says Bihorac.
Conversely, patients with a higher risk of complications would be appropriately steered to a unit with higher surveillance when needed. “The bi-directional approach is what we call optimization — avoiding the unnecessary care or procedures or tests for people who don’t need them and focusing more intensity on those who really need those interactions,” Bihorac says.
For all its potential, MySurgeryRisk is not a replacement for a human doctor. “Our goal with MySurgeryRisk and all the artificial intelligence algorithms that we’re developing is never to replace human decision-making or the doctor-patient relationship,” says Dr. Tyler Loftus, a trauma and acute care surgeon at UF and principal investigator on the AI project. Rather, the aim is to “augment it with the idea being that a doctor unaided by an algorithm is not going to perform quite as well as a doctor who benefits from having their intuition and the benefit of experience anchored in the objectivity of a highly accurate algorithm output.”
When compared to outcomes projected by surgeons evaluating 100 cases, MySurgeryRisk scored about the same as its human counterparts in predicting sepsis, prolonged mechanical ventilation, delirium and other neurological complications. The algorithm was better than humans at predicting blood clots — though Bihorac warns against getting hung up on that point.
“Our hypothesis was not superiority. We wanted to show the algorithm is not worse than humans,” she says. “Nothing can replace the human ability to observe in real time another human because there are so many visual cues that we take away from looking, smelling, understanding our experience in seeing somebody — that is unreplaceable. We wanted to make sure that even without that, that the algorithm is not inferior, and it wasn’t.”
- The Idea: An automated artificial intelligence platform that can accurately predict post-operative complications using data from electronic health records.
- Why It Matters: Better forecasting can help surgeons optimize operating-room strategies and post-operative care, as well as help with patient education. “Our vision is also to open this app for patients so you can understand what to expect — even before you come for surgery, you can start understanding as a patient where you fit into this, and how your risk points compare to others,” says Dr. Azra Bihorac, a lead researcher on the project. “The algorithm can help you prepare for meeting your doctor and asking the right questions.”
- Commercial Potential: The system has been patented and is awaiting FDA approval. “The idea really is to scale up all of this,” Bihorac says.
- The Bigger Picture: The success of MySurgeryRisk in various trials comes as the University of Florida is ramping up its AI initiatives. In 2020, the school announced a $70-million AI partnership with the California-based company Nvidia that provided UF with HiPerGator3.0, one of the 25 fastest supercomputers in the world. UF is hiring 100 AI faculty across a range of disciplines and is incorporating AI into every academic field.
- Crystal Ball: Bihorac sees AI completely transforming health. In addition to MySurgeryRisk, UF researchers are working on immersive technology known as “ambient AI,” which uses environmental sensors, cameras and other devices to monitor and interpret a patient’s condition and potentially aid doctors’ decision-making. Researchers are also working on creating virtual hospitals that will create a parallel hospital universe — or “metaverse” — where patients will be able to one day meet their care teams virtually and learn everything they need to know about their upcoming surgery without even leaving their home. “Responsible AI can allow us to actually have human-centered health care that extends beyond the world in which we are living right now,” Bihorac says.