November 3, 2022 | Diagnosing sepsis is a longstanding challenge in medicine because the clinical signs and symptoms are nonspecific and, in the absence of a gold standard for identifying patients with the disease, the focus has remained largely on identifying the culprit pathogen, says Chaz Langelier, M.D., Ph.D., associate professor of medicine in the division of infectious diseases at the University of California, San Francisco (UCSF), and a Chan Zuckerberg Biohub investigator. But an assay developed by UCSF researchers that deploys metagenomic next-generation sequencing (mNGS) of both pathogens and host immune response could be a real gamechanger.
With the aid of machine learning, the diagnostic methodology demonstrated near-perfect performance, as reported recently in Nature Microbiology (DOI: 10.1038/s41564-022-01237-2). Results included the identification of 99% of microbiologically confirmed bacterial and viral sepsis cases, and prediction of sepsis in 74% of clinically suspected cases that hadn’t been definitively diagnosed.
The research team was understandably pleased and envisions the approach being used as both a diagnostic and rule-out test in hospital emergency departments with results being returned within 48 hours, says Langelier. It could conceivably provide, “one of the most definitive assessments of sepsis status to guide precision antibiotic use,” to mitigate the existing uncertainty that exists when critically ill patients are admitted to the hospital.
“The COVID pandemic has worsened the global burden of sepsis, in part due to [increased] hospital-acquired secondary bacterial infections... and COVID itself is a leading cause of viral infections so, unfortunately, things aren’t trending in the right direction,” says Langelier. “There is an urgent need for better tests to identify patients with sepsis.”
Pre-pandemic, sepsis was calculated to account for one in five deaths globally, he adds. Among hospitalized patients who die, sepsis is the culprit in a high percentage of cases. About 30% of patients who get sepsis don’t survive.
The heterogenous nature of the disease can lead to ineffective or inappropriate treatment that isn’t matched to the pathogen causing the problem, he explains. In patients who are critically ill, “it can be very difficult to distinguish infectious from non-infectious conditions that can elicit systemic inflammatory responses [e.g., blood clots in the lungs] or other states of acute critical illness,” characterized by fevers and elevated white blood cell counts.
Most current sepsis diagnostics focus exclusively on detecting bacteria by growing them in culture, but the process is time-consuming and doesn’t always identify the bacterium that is causing the infection, says Langelier. Similarly for viruses, PCR tests don’t always identify the infectious microbe. Diagnostic tests designed to find host biomarkers associated with infection likewise have limited utility when used in isolation.
The problem is that “sepsis itself is a two-sided disease” involving a dysregulated host response and life-threatening organ dysfunction, continues Langelier. Consequently, clinicians are unable to identify the cause of sepsis in up to half of all cases.
To evaluate both sides of the problem, the UCSF team deployed mNGS to get a snapshot of the host immune response in terms of gene expression and a profile of pathogens that are in a patient’s bloodstream.
Metagenomic NGS has been used clinically for over a decade now by various academic groups and companies to detect pathogens, he says. UCSF pioneered the first use of mNGS for pathogen-seeking in cerebrospinal fluid, for example, and Karius (also located in the San Francisco Bay Area) specializes in plasma-based mNGS searching for hundreds of microbial suspects.
For their latest study, the UCSF team used mNGS to identify all the nucleic acids or genetic data present in a sample and then compared the information to reference genomes to identify the microbial organisms present—be it bacterial, viral, or fungal. Transcriptional profiling quantified gene expression to capture each patient’s response to infection. A machine learning algorithm then confirmed the diagnosis by distinguishing between sepsis and other critical illnesses.
This was a prospective, observational cohort study of more than 300 adults admitted to the intensive care unit (ICU) for a diversity of critical illnesses, “many of which at the time of [initial] emergency department admission were indistinguishable clinically in terms of infection status,” Langelier explains. “In order to identify which patients truly had an infection, we used the best adjudication approach we could think of and that was a retrospective chart review by at least three physicians following hospital discharge with all the available clinical information [blinded to the sequencing results].”
Patients were grouped into one of five sepsis categories based on their trajectory in the hospital, their standard clinical microbiology results, the treatment they received, and their outcomes, he says. The categories were for patients with clinically adjudicated sepsis and a bacterial culture-confirmed bloodstream infection; sepsis due to a microbiologically confirmed primary infection at a peripheral site other than the bloodstream; suspected sepsis with negative clinical microbiologic testing; no evidence of sepsis and a clear alternative explanation for their critical illness (e.g., pulmonary embolism, heart failure, cardiac arrest); and indeterminate status.
The machine learning model was trained on the groups where confidence in a sepsis diagnosis was the highest, and then used to predict sepsis status in the other groups, says Langelier.
For bloodstream infections, the sensitivity was 100% for the integrated host-microbe model, as reported in the published paper. For patients with microbiologically confirmed sepsis due to an infection not in the bloodstream (e.g., respiratory or urinary tract infection), the sensitivity was 97%.
Moreover, in patients with suspected sepsis but negative clinical testing, the model classified 74% as sepsis-positive and in 53% the pathogen was identified. Among patients with indeterminant sepsis, 89% were classified as having sepsis and the model identified the responsible pathogen in three out of the eight cases.
Another particularly interesting and surprising observation in the study was the ability to identify a host signature of sepsis from plasma transcriptional profiling, Langelier points out. “That has potentially important clinical implications because plasma is one of the most widely collected sample types during your average hospital admission” and the assay would therefore be broadly applicable.
The UCSF team has been working in the general space of host-microbe metagenomics for over five years now, says Langelier. Initially, mNGS was being used solely to detect pathogens in the context of respiratory infections such as pneumonia which, like sepsis, are major public health and human health challenges and involve a dynamic relationship between host and microbe.
While surveilling for pathogens that might be underlying a person’s respiratory illness, he says, “we realized that most of the sequencing data we were generating was actually not from the microbe; it was from the human. So, we decided to look at the gene expression patterns between people with respiratory infections and people without respiratory infections but who had critical respiratory illnesses due to other causes.”
One “dramatic difference” seen in the pattern of gene expressions in people with infections was immune-related genes, says Langelier, suggesting mNGS could be used not only to identify pathogens but also to determine whether someone is eliciting an immune response to the detected microbe. The team quickly moved to optimize and apply the concept to the building of a better sepsis diagnostic.
Work continues using host-microbe metagenomics for detecting lower respiratory infections, he says. A paper describing how they simultaneously interrogated the pathogen, airway microbiome, and host response to diagnose and identify pathogens in 92 critically ill adult patients with acute respiratory failure published a few years ago in Proceedings of the National Academy of Sciences (DOI: 10.1073/pnas.1809700115), and a new study is soon to be published demonstrating a proof-of-concept model for the diagnosis of such conditions in a much larger cohort of children.
“We were able to improve the accuracy of the test with an AUC [area under the receiver-operating curve] close to 0.98 for identifying lower respiratory infections in children,” he says, which currently have some children’s hospitals near capacity. For the main admission drivers—respiratory syncytial virus, flu, and rhinovirus—there often isn’t an identified microbe, just like sepsis.
The all-too-common scenario is that patients admitted to the hospital with an acute respiratory illness either receive antibiotics that are not indicated, because they in fact have a viral infection, or don’t get the appropriate antibiotics when their illness indeed has a bacterial cause, says Langelier. Antibiotic resistance is a common concern with both acute respiratory infections and sepsis.
“The key need is treating what is causing the problem, not making an educated guess based on what might be causing the problem,” Langelier says. This is typically how sepsis and pneumonia are approached currently because, with or without a definitive diagnosis, acutely ill patients need to be treated. The unintended consequence is overtreatment with antibiotics and, with it, adverse effects such as allergic reactions and kidney injury.
Getting a host-microbe metagenomics assay into widespread clinical practice will take a few years, says Langelier. The next step is to validate the model’s performance in an external cohort, and plans are already underway to do that at two hospitals. This would include an examination of the impact on antibiotic usage and the length of stays in the ICU.
A clinical trial would then need to be conducted where a test developed by a CLIA-certified laboratory at either a university or an existing clinical metagenomics company demonstrates the reproducibility of test results, he continues. The assay could thereafter be ordered by clinicians who, by virtue of the completed validation study, would have evidence of its value in guiding their medical decision-making.