A multicenter research team has demonstrated how artificial intelligence and machine learning can optimize treatment selection and administration for septic shock, a life-threatening complication that is the leading cause of hospital mortality.
The team also includes Johns Hopkins University’s Suchi Sarria, who previously developed an AI-powered early warning system to reduce sepsis mortality rates in dozens of hospitals across the country. Their results are American Medical Association Journal.
Sepsis often causes low blood pressure, which can lead to life-threatening organ damage and causes more than 270,000 deaths in the United States each year. Emergency treatment includes administering fluids and various vasopressors (drugs that constrict blood vessels) to raise the patient’s blood pressure to normal levels and restore blood and oxygen flow to organs.
“How best to individualize blood pressure treatment with different therapies remains a complex open question,” said lead author Roman Piracchio, professor of anesthesia and perioperative care at the University of California, San Francisco.
“With this kind of infrastructure, instead of doing three experiments at a time, you’re doing 1,000 experiments at a time. But you’re not doing experiments; you’re learning from existing data.”
Suchi Saria
professor of computer science
International guidelines recommend using norepinephrine, which is designed to raise blood pressure, before proceeding with vasopressin, a blood pressure-raising hormone, if a patient’s blood pressure is too low. However, septic shock is a rapidly and continuously changing condition, which complicates decisions about whether and when to start vasopressin. Additionally, vasopressin is so potent that starting it too early can cause serious side effects.
“Finding the best time to start administering vasopressin has traditionally taken millions of dollars and years, setting very specific criteria and then running clinical trials to compare those criteria to standard care, which can only test one criterion at a time,” said Salyer, professor of computer science in JHU’s Whiting School of Engineering and professor of statistics and health policy in the university’s Bloomberg School of Public Health. “As it turns out, there’s a better way, and that’s using reinforcement learning.”
Reinforcement learning is a field of machine learning in which virtual agents learn through trial and error to maximize the probability of a good outcome. The research team used electronic medical records and public datasets from more than 3,500 patients from various hospitals to train a reinforcement learning model that takes into account an individual’s blood pressure, organ failure score, and other medications they are taking to determine when to start vasopressin.
The researchers then validated the model based on unconfirmed data from approximately 11,000 additional patients, confirming the algorithm’s effectiveness and verifying that its implementation reduced in-hospital mortality.
“A significant number of patients started taking vasopressin at the exact time when our algorithm would have recommended vasopressin, if it had actually been given,” Piracchio says. “So, by using complex statistical techniques to account for baseline bias and differences, we were able to show that treatments that match exactly what the algorithm suggested – treatments that start at exactly the right time – are consistently associated with better outcomes in terms of mortality.”
The model consistently recommended starting vasopressin earlier than most physicians actually do, but in the small number of cases where the drug was given sooner than the algorithm recommended, patient outcomes were worse.
“This shows that it is beneficial to try to individualize strategies for each patient,” Piracchio says. “In septic shock, there is wide variation in resuscitation practices across hospitals and countries, especially when it comes to vasopressor support. Given the heterogeneity of the population included in this study, the results demonstrate that individualized vasopressin initiation rules can improve outcomes for patients with septic shock.”
The next step is to actually implement the “promise-to-reality” model, as Sarria calls it.
Piracchio and his team plan to do just that at UCSF Medical Center before partnering with Bayesian Health, a clinical AI platform spun out of Thalia’s research, to expand to centers across the country. However, the application of reinforcement learning in medicine goes beyond administering vasopressors.
“With this kind of infrastructure, instead of running three experiments at a time, you’re running 1,000 experiments at a time. But you’re not running experiments; you’re learning from existing data,” Salyer says. “It’s as if the experiment is already being done for free, and we can learn from it and wisely discover the exact circumstances in which different strategies should be implemented to improve patient outcomes and save lives.
“There are a lot of opportunities for reinforcement learning here. This is just the beginning.”
