A machine learning tool analyzed EHR big data to determine four subtypes of sepsis, which could lead to better treatment of the condition.
By Jessica Kent
May 21, 2019 – Using machine learning to analyze big data in the EHR revealed four subtypes of sepsis, showing that sepsis patients could benefit from different types of treatment instead of a one-size-fits-all approach, according to a study published in JAMA.
Sepsis is a deadly response to infection that affects millions each year. Although researchers’ understanding of the condition has advanced considerably, studies haven’t translated into new treatments.
“A major barrier to progress is the overly broad definition of the syndrome, which encompasses a vast, multidimensional array of clinical and biological features,” the team said.
“Different combinations of these features may naturally cluster into previously undescribed subsets or phenotypes that may have different risks for a poor outcome and may respond differently to treatments.”
Researchers from the University of Pittsburgh School of Medicine used a machine learning algorithm to analyze 29 clinical variables the EHRs of over 20,000 UPMC patients recognized to have sepsis within six hours of arriving at the hospital from 2010 to 2012.
The algorithm grouped patients into four distinct sepsis types. Alpha was the most common type, with 33 percent of patients falling into this category. These patients had the fewest abnormal laboratory test results, the least organ dysfunction, and the lowest in-hospital death rate at two percent.
Patients classified as having the beta sepsis type were older and had the most chronic illnesses and kidney dysfunction. Twenty-seven percent of patients comprised this category, and the gamma sepsis type contained a similar percentage of patients as the beta type (based on elevated measures of inflammation and pulmonary dysfunction).
Delta sepsis types were the least common, with just 13 percent of patients falling into this category. This type was also the deadliest, often combining with liver dysfunction and shock, with an in-hospital death rate of 32 percent.
Researchers also used the algorithm for another 43,000 sepsis patients at UPMC from 2013 to 2014, and the results showed a similar frequency of sepsis types and clinical characteristics as the primary analysis. The same happened when the group studied clinical data and immune response biomarkers from nearly 500 pneumonia patients enrolled at 28 hospitals across the US.
These findings show that the algorithm can be applied to clinical data in multiple patient populations and yield the same results.
“These sepsis phenotypes can be identified at the time of patient presentation to the emergency department, and thus could be useful with regard to early treatment and enrollment in clinical trials,” the team said.
“Only routinely available data were used in the clustering models, and the phenotypes were derived from a large observational cohort to ensure generalizability.”
In addition to testing the algorithm across several datasets, the team applied the tool to data from several recently completed international clinical trials. The trials had tested different promising sepsis treatments, and they had all ended with average results.
With trial participants sorted into the four sepsis subtypes, the researchers found that some clinical trials could have had better results. For example, a 2014 study on early goal-directed therapy (EGDT), an aggressive resuscitation protocol that includes placing a catheter to measure blood pressure and oxygen levels, was found to have no benefit.
However, after the researchers re-evaluated results and used the machine learning algorithm to classify the participants, they found that EGDT was beneficial for alpha-type sepsis patients. On the other hand, EGDT was found to have detrimental effects on delta-type patients, an observation that demonstrates the critical role the subtypes could play in future sepsis treatment research.
“These proof-of-concept clinical phenotypes could be incorporated prospectively in future study designs that test new biologically active therapeutics. Novel designs could enrich for a priori phenotypes as well as confirm the boundaries around predictive phenotypes during the trial,” the researchers said.
The team noted limitations of the study given that the characteristics of the phenotypes came from a single health system in the US. While the phenotypes were found to be generalizable in other datasets, the group said that further research is necessary.
Despite this and other limitations, researchers are confident that these four classifications of sepsis will help improve treatment for each patient who develops the condition.
“Intuitively, this makes sense—you wouldn’t give all breast cancer patients the same treatment. Some breast cancers are more invasive and must be treated aggressively. Some are positive or negative for different biomarkers and respond to different medications,” said senior author Derek Angus, MD, MDH, professor and chair of Pitt’s Department of Critical Care Medicine.
“The next step 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.”