You are here
June 4, 2019
Machine learning reveals four types of sepsis
At a Glance
- Researchers used machine learning to identify four types of sepsis that may affect treatment outcomes.
- The findings could aid the development of better treatment strategies.
Sepsis is the body’s extreme response to an infection. The infection triggers a chain reaction throughout the body. It causes fever, chills, rapid breathing, confusion, and extreme pain or discomfort. Without timely treatment, it can quickly lead to organ failure and death. Each year, more than 1.7 million people nationwide get sepsis, and about 270,000 die from it.
Doctors treat sepsis by trying to cure the infection while supporting vital functions with such things as oxygen and intravenous fluids. Previous studies have found that people with sepsis have different infections, symptoms, lab test results, and treatment outcomes. A team of researchers led by Drs. Derek C. Angus and Christopher W. Seymour at the University of Pittsburgh set out to define different types of sepsis by using large collections of health record data. The study, which was funded primarily by NIH’s National Institute of General Medical Sciences (NIGMS), appeared in JAMA on May 19, 2019.
First, the team used a machine learning approach to analyze 29 variables, such as age, blood pressure, and lab test results, from the electronic health records of more than 20,000 people. These patients were diagnosed with sepsis within 6 hours of arrival at a system of 12 hospitals in Pittsburgh in 2010-2012.
The machine learning approach looked for patterns in the patient health variables. It determined that the data could be clustered into four types of sepsis. The most common type (33% of patients) was called alpha. These patients had the most normal lab test results and lowest death rate. The beta type (27% of patients) consisted of older people who tended to have chronic illness, especially kidney problems. The gamma type (27%) had more inflammation, higher fever, and breathing problems. The delta type (13%) was associated with liver problems and dangerously low blood pressure. People with the delta type were most likely to die.
The team tested the machine learning approach on a second set of electronic health records for more than 43,000 sepsis patients treated in 2013-2014. Again, they found four types of sepsis with similar clinical characteristics. The proportions for the four types were also similar to the 2010-2012 results: 29% for alpha, 29% for beta, 28% for gamma, and 14% for delta.
The researchers next analyzed results from several clinical trials. They found that accounting for the different types of sepsis could have affected the outcomes of these trials. This finding shows that sepsis types need to be considered in the design of clinical trials.
“Hopefully, by seeing sepsis as several distinct conditions with varying clinical characteristics, we can discover and test therapies precisely tailored to the type of sepsis each patient has,” Seymour says.
“The next step,” Angus adds, “is to do the same for sepsis that we have for cancer—find therapies that apply to the specific types of sepsis and then design new clinical trials to test them.”
—by Geri Piazza
References: Derivation, Validation, and Potential Treatment Implications of Novel Clinical Phenotypes for Sepsis. Seymour CW, Kennedy JN, Wang S, Chang CH, Elliott CF, Xu Z, Berry S, Clermont G, Cooper G, Gomez H, Huang DT, Kellum JA, Mi Q, Opal SM, Talisa V, van der Poll T, Visweswaran S, Vodovotz Y, Weiss JC, Yealy DM, Yende S, Angus DC. JAMA. 2019 May 19. doi: 10.1001/jama.2019.5791. [Epub ahead of print]. PMID: 31104070.
Funding: NIH’s National Institute of General Medical Sciences (NIGMS) and GlaxoSmithKline.