Northumbria Healthcare NHS Foundation Trust adopts the Responsible AI philosophy with Azure Machine Learning

Understanding the risks of surgery

Orthopedic surgeons Mike Reed and Justin Green of Northumbria Healthcare NHS Foundation Trust know that the possibility of complications is an important consideration in any joint-replacement procedure—the doctor, the patient, and the hospital all want to understand and mitigate potential problems. With better information at hand, patients can make a more-informed decision about whether to have a procedure, doctors and surgical staff can be on the alert for possible issues, and hospital schedulers can assign patients to a facility with a level of care appropriate to their risk factors.

“When a patient looks one of us in the eye and asks if the surgery will go all right, we often end up giving quite general answers because it can be very hard to predict,” says Reed. “That led us to explore how technology can give us a better indication of how patients will fare so we can have more meaningful conversations with them and better plan for their surgery.”

Northumbria Healthcare NHS Foundation Trust serves a population of approximately 500,000 people in northeast England across one of the largest geographic regions of any NHS trust. Like healthcare facilities around the world, Northumbria Healthcare NHS Foundation Trust developed a backlog of surgery candidates during the COVID-19 pandemic, as elective procedures were put on hold to free up space for critical care. Fortunately, Reed and Green had already begun explorations into technology that had begun to produce results that can help triage these patients to the right doctors and operating facilities.

Turning data into medical insights

In addition to his surgical responsibilities, Reed is an enthusiastic researcher who has done extensive studies on patient outcome data from the NHS to determine the most significant surgical risk factors. After analyzing very large data sets of as many as 400,000 patients using traditional statistics, he began to consider other methods. “A few years ago, I asked some university students to explore how AI and machine learning might enhance our research,” says Reed. “Within weeks, students produced some quite dramatic data that seemed much better than our existing risk-assessment protocols.”

Reed was excited about the possibility of developing a full production system to draw further insights from the data, and he got funding from several NHS organizations in late 2021. He also connected with Microsoft. “Microsoft gave us valuable data science advice that really transformed how fast we were able to work,” says Reed. “Our team is small and mostly clinically focused—we’re not professional data scientists, but with help from Microsoft we’ve added strong data science skills to our toolkit.”

Reed and his team also began using Microsoft Azure Machine Learning for the project. In addition to its powerful machine learning capabilities, the service also provides Reed and his team access to the Responsible AI dashboard. This comprehensive one-stop interface brings together several mature Responsible AI tools in the areas of machine learning interpretability, unfairness assessment and mitigation, error analysis, causal inference, data exploration, and counterfactual analysis. These capabilities were essential for the development of a robust and clinically adoptable healthcare AI application.

“In healthcare, AI can’t just be a black box that takes inputs, performs unknown calculations on them, and produces a result,” says Green. “If we’re going to change a patient’s care plan based on the results of the application, we need to know why it generated that result, and we need to be sure that the result isn’t affected by demographic bias or other factors. With Azure Machine Learning and the Responsible AI dashboard, we have the tools to understand, refine, and explain our outcomes so we can better serve our patients.”

The fact that the cloud and AI solutions came from Microsoft was important to the project team. “Credibility is a key issue, and I know that Microsoft hires the best people and keeps all its cloud technologies very up to date,” says Reed. “I think it’s fair to say that Microsoft has more experience scaling software than anybody else, so that gave us reassurance that we would be able to expand our solution beyond our one hospital to the rest of the country and even the rest of the world.” To facilitate adoption by other healthcare organizations, the team is making its models and methodologies available as open-source software. 

Generating unexpected results and personalizing healthcare

For the past two-and-a-half years, Green has been part of Reed’s team, where he gets to combine his career in surgery with his interest in digital health and informatics. With his help, the team has uncovered some interesting results from the 220 different input parameters its machine learning model uses.

“We know that factors like age, ethnicity, smoking history, and body mass index are important determinants for surgical complications,” he explains. “But after we added blood test results into our machine learning models, we discovered that platelet count was one of the most important predictive factors for the success of an operation. This was really surprising and as we were able to distinguish between the level of influence with such granularity, something we hadn’t been able to do before.”

Having this additional information can make for deeper patient interactions. “People may think of AI and machine learning as depersonalizing technologies, but in reality, they can make the delivery of healthcare much more personal,” says Green. “They also help us, as clinicians, understand just how unique people are, and this guides us away from a one-size-fits-all risk model and toward something much more individualized.”

Building on success and delivering better care

Reed, Green, and their team will fully scale out the deployment within Northumbria in the coming months, but they are already excited to expand the project’s scope. The current application is focused on hip and knee replacements, but it has become clear to the team that the work is applicable to a wide range of surgeries and other medical procedures. “We can use the same structured approach and the same techniques to explore questions for other use cases,” says Green. “And with the Responsible AI dashboard, we can conduct our inquiries in an explainable, unbiased way across a myriad of cases.”

While the team starts working to help clear the COVID surgical backlog in Northumbria by better assessing people in line for procedures, it is also expanding its efforts beyond the Northumbria Trust. “We are collaborating with an NHS hospital group in Wales to design a pilot program to help assess their surgical waiting list,” says Reed. “They have one hospital with 10,000 patients waiting, and our work can help identify patients who could safely have surgery at a lower-specification hospital, thereby reducing the wait for both lower and higher-risk patients.”

While the doctors with whom the project team has shared its findings have been excited—even amazed—by the possibilities they present, it’s ultimately patients who have the most to gain. “With Azure Machine Learning, we can show the patient a risk score that is highly tailored to their individual circumstances,” explains Reed. “At that point, they can decide whether they want to have the surgery at all, or if there are lifestyle changes they can make to decrease their risk before going forward. Ultimately, we aim to reduce risk, reduce uncertainty, and improve surgical outcomes.”

Find out more about Northumbria Healthcare NHS Foundation Trust on Twitter and LinkedIn.