Produced in partnership with TEDMED

AI and machine learning could halve preventable errors in medicine

Researcher Suchi Saria works to bridge the gap between AI solutions and implementation in healthcare

Imagine that a friend or loved one is in the hospital for a routine operation or procedure. It seems like they are on their way to recovery, but then, you learn that septic shock — a widespread infection leading to organ failure and low blood pressure — has set in. Your loved one needs immediate and urgent medical attention. Such an outcome is not out of the ordinary. According to the CDC, about 270,000 of the 1.7 million American adults that develop sepsis will die each year, and 1 in 3 hospital deaths are the result of sepsis.

Now imagine a new scenario: earlier in your loved one's stay, hospital staff receive a notification from an early warning system that alerts then of the sepsis risk, maybe even before nurses or doctors can diagnose it. Your loved one receives life saving antibiotics right away. Such an outcome is no longer the stuff of sci-fi. Some hospital systems are already test-driving such early warning systems.

This latter scenario illustrates the potential power of incorporating artificial intelligence (AI) and machine learning (ML) into healthcare for the purposes of averting otherwise preventable deaths. This strategic application of AI to healthcare is the long-term mission of Suchi Saria, director of the Machine Learning and Healthcare Lab at Johns Hopkins University and founder and CEO of Bayesian Health.

Saria recounts that her interest in AI and machine learning started in her childhood with an intrinsic drive to understand how things worked. "I was always a tinkerer," Saria admitted. Her interest in programming, building robots, and ultimately, trying to understand what it meant to build intelligent machines, placed her in the forefront of research in AI and machine learning.

A person typing on a laptop with a stethoscope next to them

National Cancer Institute via Unsplash 

However, while the problems that she was working on were exciting in a technical sense, she started to question their social impact. So, in the late 2000s, Saria began collaborating with doctors at Stanford University working with premature babies and had the exciting realization that data pulled from EMR systems could help to identify premature babies at high risk of complications. "We were able to show that it's possible to analyze physiologic data and find patterns of deterioration associated with different complications in babies, and these patterns were detected…much earlier than was visible to the human eye," Saria said.

Even with these inroads, Saria faced a personal reckoning with the loss of her nephew from sepsis in 2014. Saria expressed her frustration that despite writing several scientific papers that, "none of that was actually leading to improved outcomes in any way." For Saria, her company, Bayesian Health, "was really built to meet this unmet need that today…we've invested billions of dollars in collecting this data, but the use of this data in improving outcomes is still very limited." However, the goal of integrating these tools into healthcare remains a non-trivial task. In her 2020 TEDMED talk, Saria outlined three main barriers to implementation. First, challenges remain with integration and access to electronic medical record (EMR) data. Secondly, the incentives for hospitals are mismatched: hospitals aren't financially rewarded for saving lives, they are paid for doing more procedures. Finally, it is necessary to build more trust among physicians. "The quality of the machine learning necessary …for providers to trust, it is a very high bar," Saria explained.

X-rayed hands showing the "OK" sign

Modified from Owen Beard on Unsplash

Saria predicts that AI will find applications across the board in health care ranging from specialty care models, targeting expensive therapies, or "stitch[ing] together data across different settings in order to optimally care for patients." The exciting news according to Saria is that the infrastructure, data, and ability to deliver information to workflow providers already exists: "I think we have the right solutions. Now, it's a question of dissemination, distribution." Ultimately, Saria is optimistic about the impact that AI can have in healthcare. She projects that in the next five years we could reduce preventable harms and diagnostic errors by 50%.