Penn leverages machine learning to identify severe sepsis early
Researchers have developed an algorithm that predicts septic shock in hospitalized patients using data from an electronic health record.
A machine learning algorithm has been developed by researchers at the University of Pennsylvania Health System to identify hospitalized patients most at risk for severe sepsis or septic shock by leveraging electronic health records, and then using the EHR to alert care teams.
“One of the major causes of preventable mortality in hospitals is sepsis. And, there’s some evidence that sepsis is not recognized as early as it could be,” says Craig Umscheid, MD, associate professor of medicine and epidemiology at the University of Pennsylvania Perelman School of Medicine. “There’s also good evidence that the sooner you recognize sepsis and initiate effective treatment, the better the outcome.”
Delayed diagnosis and treatment of sepsis, an immune system response to infection, has made it one of the most deadly and costly medical conditions confronting U.S. hospitals, with more than 1 million cases each year and 250,000 fatalities annually.
However, last week researchers from the University of Pennsylvania Health System presented results of their study at the 2017 American Thoracic Society International Conference, showing that the machine learning algorithm has the capability to continuously sample real-time EHR data to prospectively identify patients at risk for severe sepsis and septic shock, enabling additional monitoring and early intervention.
“We have developed and validated the first machine learning algorithm to predict severe sepsis and septic shock in a large academic multi-hospital healthcare system,” said lead author Heather Giannini, MD, internal medicine resident at the Hospital of the University of Pennsylvania. “This is a breakthrough in the use of machine learning technology, and could change the paradigm in early intervention in sepsis.”
Giannini points out that diagnosing sepsis is particularly difficult because it is a syndrome that lacks a gold standard clinical test for diagnosis.
“People can look as though they have sepsis without the infection,” she says. “Their abnormalities, vitals, labs, and organ failures can be due to something else. Because there is such a high mortality rate associated with sepsis, there is no room for error.”
Also See: HDM’s Ingite:Digital Health conference, July 13-14
According to Giannini, there are definitive tell-tale signs for sepsis “but by the time those develop, you are already at a point where the patient is at very high risk for mortality.” As a result, she makes the case that monitoring and analyzing real-time EHR data is critical to early identification and intervention.
“The signals are there to show us that there is very high likelihood of clinical deterioration due to the development of sepsis prior to what we would traditionally be looking for,” observes Giannini. “The power of machine learning lies in the ability to analyze enormous amounts of data and to use these decision trees” as a decision support tool producing “more robust” predictions.
The key is “having the electronic health record and the tools we’ve built from data in the EHR tell us something new, rather than just alerting us to something that we already know,” adds Umscheid.
Also See: Alert system uses predictive analytics to combat sepsis
The sepsis machine learning algorithm was developed internally by researchers who trained a random forest classifier—an approach to classify a wide range of data—to sort through EHR data such as labs, vitals and demographics for 162,212 patients discharged between July 2011 and June 2014 from three University of Pennsylvania Health System acute care hospitals. The algorithm was then validated in real time between October and December 2015 with 10,448 patients while they were cared for in the study hospitals, using a “silent mode” of EHR sampling.
A total of 943 patients in the training set met the lab or physiological criteria and were coded with either severe sepsis or septic shock. Sensitivity and specificity for identification of patients who were coded with severe sepsis or septic shock were 26 percent and 98 percent, respectively. Positive and negative predictive values were 29 percent and 97 percent, respectively, while the likelihood ratios—positive and negative—were 13 and 0.75, respectively.
The algorithm and the EHR alert for sepsis were implemented on the Allscripts Sunrise Clinical Manager, according to Umscheid. However, he notes that the University of Pennsylvania Health System recently switched its inpatient EHR system from Allscripts to Epic. “We will be making decisions about what the next version looks like in our Epic EHR.”
“We have not made the transition over to Epic just yet” for the algorithm and alert, adds Giannini. “Our data scientists are early in the process working with them in order to transition things over. It’s a work in progress.”
Going forward, she says researchers will make a more extensive assessment of how the machine-learning algorithm has impacted the University of Pennsylvania Health System and provider teams as well as evaluate the algorithm’s performance in more real-time clinical settings over a longer study period.
Ultimately, the success of the algorithm will be judged on whether its predictions improve the implementation of best practices for patients and clinical outcomes, concludes Umscheid.
“One of the major causes of preventable mortality in hospitals is sepsis. And, there’s some evidence that sepsis is not recognized as early as it could be,” says Craig Umscheid, MD, associate professor of medicine and epidemiology at the University of Pennsylvania Perelman School of Medicine. “There’s also good evidence that the sooner you recognize sepsis and initiate effective treatment, the better the outcome.”
Delayed diagnosis and treatment of sepsis, an immune system response to infection, has made it one of the most deadly and costly medical conditions confronting U.S. hospitals, with more than 1 million cases each year and 250,000 fatalities annually.
However, last week researchers from the University of Pennsylvania Health System presented results of their study at the 2017 American Thoracic Society International Conference, showing that the machine learning algorithm has the capability to continuously sample real-time EHR data to prospectively identify patients at risk for severe sepsis and septic shock, enabling additional monitoring and early intervention.
“We have developed and validated the first machine learning algorithm to predict severe sepsis and septic shock in a large academic multi-hospital healthcare system,” said lead author Heather Giannini, MD, internal medicine resident at the Hospital of the University of Pennsylvania. “This is a breakthrough in the use of machine learning technology, and could change the paradigm in early intervention in sepsis.” Giannini points out that diagnosing sepsis is particularly difficult because it is a syndrome that lacks a gold standard clinical test for diagnosis.
“People can look as though they have sepsis without the infection,” she says. “Their abnormalities, vitals, labs, and organ failures can be due to something else. Because there is such a high mortality rate associated with sepsis, there is no room for error.”
Also See: HDM’s Ingite:Digital Health conference, July 13-14
According to Giannini, there are definitive tell-tale signs for sepsis “but by the time those develop, you are already at a point where the patient is at very high risk for mortality.” As a result, she makes the case that monitoring and analyzing real-time EHR data is critical to early identification and intervention.
“The signals are there to show us that there is very high likelihood of clinical deterioration due to the development of sepsis prior to what we would traditionally be looking for,” observes Giannini. “The power of machine learning lies in the ability to analyze enormous amounts of data and to use these decision trees” as a decision support tool producing “more robust” predictions.
The key is “having the electronic health record and the tools we’ve built from data in the EHR tell us something new, rather than just alerting us to something that we already know,” adds Umscheid.
Also See: Alert system uses predictive analytics to combat sepsis
The sepsis machine learning algorithm was developed internally by researchers who trained a random forest classifier—an approach to classify a wide range of data—to sort through EHR data such as labs, vitals and demographics for 162,212 patients discharged between July 2011 and June 2014 from three University of Pennsylvania Health System acute care hospitals. The algorithm was then validated in real time between October and December 2015 with 10,448 patients while they were cared for in the study hospitals, using a “silent mode” of EHR sampling.
A total of 943 patients in the training set met the lab or physiological criteria and were coded with either severe sepsis or septic shock. Sensitivity and specificity for identification of patients who were coded with severe sepsis or septic shock were 26 percent and 98 percent, respectively. Positive and negative predictive values were 29 percent and 97 percent, respectively, while the likelihood ratios—positive and negative—were 13 and 0.75, respectively.
The algorithm and the EHR alert for sepsis were implemented on the Allscripts Sunrise Clinical Manager, according to Umscheid. However, he notes that the University of Pennsylvania Health System recently switched its inpatient EHR system from Allscripts to Epic. “We will be making decisions about what the next version looks like in our Epic EHR.”
“We have not made the transition over to Epic just yet” for the algorithm and alert, adds Giannini. “Our data scientists are early in the process working with them in order to transition things over. It’s a work in progress.”
Going forward, she says researchers will make a more extensive assessment of how the machine-learning algorithm has impacted the University of Pennsylvania Health System and provider teams as well as evaluate the algorithm’s performance in more real-time clinical settings over a longer study period.
Ultimately, the success of the algorithm will be judged on whether its predictions improve the implementation of best practices for patients and clinical outcomes, concludes Umscheid.
More for you
Loading data for hdm_tax_topic #better-outcomes...