Listen to a discussion of this article through our AI-generated podcast.
Beginning with our September 2024 issue, CAP TODAY now offers the option to listen to a discussion of our articles with voices reminiscent of a Siri or Garmin speaker. Since this AI-generated podcast uses technology that is still in its early stages of development, there may be some errors in pronunciation and the tone may sound too upbeat for the material covered.
Amy Carpenter
September 2024—When emergency physician Chadd Kraus, DO, DrPH, of Lehigh Valley Health Network in Allentown, Pa., sees a patient who could have sepsis, he wants to know if there’s an infection and, if so, how bad it is based on the patient’s host response, what interventions are needed, and whether the patient will need to be hospitalized. Not as pressing for him? “What’s the pathogen?”
What Dr. Kraus wants in a “good test,” he said, is something that will guide him to the “next best step,” not necessarily the definitive answer. “And that’s sometimes a perspective unique to the ED,” he said.
He doesn’t have to know if it’s bacterial, viral, or fungal. “I need to know that this patient is really sick and I have to get things started.” In the first 10 minutes, he doesn’t need to know the pathogen, he said, but “it would be great to know if it’s sepsis.”
He presented with Alison Woodworth, PhD, D(ABCC), and Nathan Ledeboer, PhD, D(ABMM), this summer in an ADLM session on “Pathogen Detection, Machine Learning, and Host Biomarkers: Can Any of These Technologies Help Address the Challenge of Sepsis?”
Dr. Ledeboer, professor of pathology and chief, Division of Clinical Pathology, Department of Pathology and Laboratory Medicine, Medical College of Wisconsin, provided a close-up of the current or near-market solutions for host response detection. Next month we’ll go back to the sepsis basics with Dr. Woodworth of CTI Clinical Trial and Consulting.
Drs. Ledeboer and Kraus and others studied the consistency of provider judgment and the potential of a new host response sepsis test to aid in the triage process, using case vignettes (Kraus CK, et al. J Pers Med. 2023;13[12]:1685).
Their modified Delphi study involving 26 participants from multiple specialties found provider assessment of sepsis risk in the cases to range from 10 percent to 90 percent and agreement was poor.
The rapid host response test they chose to center the research around was the Cytovale IntelliSep test, which the FDA cleared in 2022. “It basically compares WBCs from healthy patients to those of septic patients,” Dr. Ledeboer explained, “and there is a substantial amount of deformation that occurs in a septic patient compared to a healthy donor.” He describes it as a simple and rapid method that “uses an approach similar to flow cytometry that many of us are used to. And the test can do this over thousands and thousands of measurements in a very short time.”
The measurements from the test are then fed into an algorithm to classify a patient based on how cells deform, putting patients into three groups and providing a clear indication to the ED clinician when it’s needed most. Three interpretation bands are based on the probability of sepsis with organ dysfunction manifesting within the first three days after testing and tell the clinician the patient is unlikely to be septic, is a cautionary patient (slow down, additional workup may be appropriate), or one who needs aggressive and intensive management.
“This helps us to meet Dr. Kraus’ concern,” said Dr. Ledeboer, who is also associate chief medical laboratory officer, Froedtert Health, Milwaukee.
In another study of the IntelliSep test, Dr. Kraus and others evaluated its potential to expedite appropriate care for patients who present with signs of infection (O’Neal HR Jr, et al. Acad Emerg Med. Published online April 21, 2024. doi:10.1111/acem.14923). They performed a pooled analysis of five adult cohorts at seven U.S. sites in separate studies. Structured blinded adjudication was used to classify presence or absence of sepsis.
Patients classified by the test as band three (high risk) had a very high likelihood of sepsis by the clinical adjudicators—“a better than 50 percent chance,” Dr. Ledeboer said. Band two (cautionary group) had an elevated chance of sepsis of about 25 percent. “And band one—almost zero of the patients had sepsis.”
The authors wrote, “If integrated into standard of care, the test may help improve outcomes and reduce unnecessary antibiotic use.”
Other approaches such as transcriptomics or proteomic-based technologies “may give us a bit better specificity,” Dr. Ledeboer said, “because we can look at multiple markers in a single assay and account for a bit more of the geographical variation that can occur.”
He cited a study that evaluated transcriptomics, defined as the analysis of host gene expression through RNA transcripts, with a microarray-based approach using peripheral blood (Holcomb ZE, et al. J Clin Microbiol. 2017;55[2]:360–368). “It then can identify a number of classifiers to evaluate the patient populations,” Dr. Ledeboer said. Transcriptomics can quickly provide multiple data points: whether the infection is bacterial or viral and its severity.
Two commercial assays use a four-point scale—very unlikely, unlikely, possible, or very likely—to assign sepsis risk, he said.
Data from the SeptiCyte Rapid assay (Immunexpress) 510(k) submission reveal the probability of sepsis with a score broken into four interpretive bands: forced, consensus, unanimous, or indeterminate adjudication. “We saw very few forced adjudications—most were in the consensus or unanimous populations,” with the probability of sepsis being well under 10 percent, Dr. Ledeboer said. In band four, the group at highest risk, “we see an 80 percent risk or probability of sepsis.”
The TriVerity Acute Infection and Sepsis Test from Inflammatix is a second transcriptomics-based approach. Here, Dr. Ledeboer cites an evaluation of the whole blood, multiplex host-messenger RNA expression metric, Inflammatix-Severity-2 (since renamed TriVerity), for identifying septic, hospitalized patients’ likelihood of 30-day mortality, development of chronic critical illness, discharge disposition, and/or secondary infections (Brakenridge SC, et al. Crit Care Explor. 2021;3[10]:e0554). The authors compared the Inflammatix-Severity-2 outcome variables with those of commonly used sepsis biomarkers, such as C-reactive protein, absolute lymphocyte counts, total WBC counts, and interleukin-6, and found the Inflammatix-Severity-2 severity score superior for predicting secondary infections and overall adverse clinical outcomes.
Dr. Ledeboer describes the third approach as AI-driven decision support technologies, whereby metadata is fed into a machine learning algorithm to help the algorithm identify several classifiers that may work clinically to predict risk of sepsis. “We subsequently evaluate the performance of the algorithm using the model developed with the training set,” he said.
He found in an online search 111 algorithms to predict sepsis risk. “Not all algorithms are created equal, and they use very different design approaches.” Some are as simple as best practice advisories, firing based only upon a clinical set of criteria. Others are adaptive, using machine learning or deep learning-based approaches to change and improve over time.
Most such algorithms, he said, were developed using existing medical record data, often based on a single center. “Most of this is going to be very structured data, but understanding whether it’s structured versus unstructured data and how that is incorporated into the algorithm is important.” Goh, et al., developed an AI algorithm using both structured data and unstructured clinical notes to predict and diagnose sepsis (Goh KH, et al. Nat Commun. 2021;12[1]:711). Mining the clinical notes, they wrote, improved the algorithm’s accuracy compared with using only clinical measures for early warning 12 to 48 hours before sepsis onset.
In a before-and-after quasi-experimental study published earlier this year, Boussina, et al., evaluated the impact on patient outcomes of a real-time, deep-learning model called Composer for the early prediction of sepsis (Boussina A, et al. NPJ Digit Med. 2024;7[1]:14). “It was that most rudimentary approach but it was supported by deep learning,” Dr. Ledeboer noted. They looked at 6,217 patients (5,065 in a pre-intervention phase, 1,152 post-intervention) who met the Sepsis-3 consensus definition at two EDs in the UC San Diego Health system.
“They took data from Epic and fed it into the Composer [algorithm] microservice, hosted on Amazon Web Services,” Dr. Ledeboer said, and this was combined with laboratory, pharmacy, and outcome data. This led to a nurse-facing best practice advisory in the patient’s chart triggered by Composer with alerts identifying the patient’s risk of SIRS, sepsis, severe sepsis, or septic shock. The authors reported about a 13 percent reduction in sepsis-related mortality. “They saw about a five percent increase in the qSOFA 72 score, which indicated risk of organ failure decreased,” Dr. Ledeboer said.
“All of these potential approaches have merit,” he said. “We just need to see a bit more data on each of these approaches over time.”
A study of the impact of host response on length of stay and antibiotic use pre- and post-implementation of a rapid sepsis diagnostic approach (IntelliSep, post-FDA clearance) was presented this spring (Thomas CB, et al. Poster P1213 presented at European Society of Clinical Microbiology and Infectious Diseases; April 27–30, 2024; Barcelona).
The authors reported an average two-day decrease in length of stay among survivors, Dr. Ledeboer noted. “They also observed an antimicrobial stewardship benefit from about 20 antibiotic doses per 100 ED visits to around 16 antibiotic doses per 100 ED visits” between July 1 and Oct. 31, 2023.
The authors found, too, a reduction in blood culture orders, from about 60 percent of patients in Aug. 1, 2023 to about 40 percent of patients by Oct. 31, 2023. “Patients who were at the highest risk of sepsis still got blood cultures” when they were needed, Dr. Ledeboer said. For patients at the lowest risk of sepsis, blood culture orders declined from 40 percent to less than 20 percent. “So from a laboratory stewardship standpoint, this is a great outcome.”
“As we start to think more about use of these host response assays,” Dr. Ledeboer said, “it will enable us to be a bit smarter about how we use direct pathogen detection,” which has “largely been a failure in patients with bloodstream infections because it’s expensive, and we’re asking it to identify bacteria in patients where we see a less than 10 percent positivity rate.”
The T2Bacteria panel of T2 Biosystems was evaluated in a prospective multicenter study of 1,427 patients for whom blood cultures were ordered as standard of care (Nguyen MH, et al. Ann Intern Med. 2019;170[12]:845–852). The authors reported that the assay accurately identified or excluded bloodstream infections caused by five common pathogens in about four to eight hours versus about 24 to 72 hours and five days, respectively, for blood cultures.
“It’s expensive,” Dr. Ledeboer said. “It costs $100 to $250 to run a direct pathogen detection test. So if we’re using it for a screening, it becomes a different argument. But if it can enrich those patients who get a rapid pathogen detection test, we can deploy those resources to where they’re most needed, those highest-risk patients.”
Why is host response so impactful? Dr. Ledeboer asks. It’s because patients at the highest risk of sepsis can be immediately started on the appropriate anti-infectives and the appropriate resuscitation can be initiated. With patients at moderate risk, a clinical exam with laboratory data can be used to further evaluate sepsis risk, and “patients at the lowest risk get better care because instead of wasting time chasing the risk of sepsis, we can evaluate them for other causes of shock,” such as cardiogenic, hypovolemic, or obstructive shock. “We can defer microbiology testing.”
By using a host response approach, he argues, “we can be a bit smarter when it comes to direct pathogen detection.”
Amy Carpenter is CAP TODAY senior editor.