Sepsis, a life-threatening organ dysfunction, is responsible for an estimated one in five deaths globally and one in three deaths that occur within a U.S. hospital. It happens due to a dysregulated immune response to infection, and its high mortality rate correlates to the difficulty in quickly diagnosing and appropriately addressing the origin of infection, be it bacterial, viral, or some other pathogen. Current methods can delay that life-or-death determination for as long as three days.

Enter the Stanford Medicine lab of Purvesh Khatri, where a test developed to reduce that time frame to roughly 30 minutes – an advancement that could be transformative for millions of patients each year – was cleared by the Food and Drug Administration in January.

Purvesh Khatri | Courtesy Stanford Medicine

Khatri, PhD, a faculty member in the Institute for Immunity, Transplantation and Infection and the Department of Medicine, says if a patient's sepsis is caused by a bacterial infection, every hour that passes before antibiotics are administered increases the risk of mortality by 6-8%. As a result, emergency department physicians often prescribe antibiotics even before infection has been identified, just in case it is bacterial.

If the real culprit is a virus, though, the antibiotic treatment can be counterproductive as it may create a favorable environment for antibiotic-resistant bacteria. And if the cause is sterile inflammation, neither bacterial nor viral but something else – related to cancer, surgery, blood clots, or internal injury, for example – delay in treatment can also be catastrophic.

“Time is of the essence in the proper diagnosis of sepsis,” says Khatri. ”Emergency department physicians have been hobbled by the need to wait so long to identify a patient's infection as viral or bacterial. It’s kept them from rapidly being able to apply the most effective and appropriate treatment with confidence.”

In 2016, Khatri, Tim Sweeney, MD, PhD, and Jonathan Romanowsky, co-founded a company, Inflammatix, to develop a platform for the sepsis test. Called TriVerity™, The test evaluates a blood sample for 29 genes that are implicated in the presence, type, and severity of infection. The gene signatures that constitute the test's foundation were identified by Khatri at Stanford.

The recent FDA clearance marks not only an important step toward the realization of a potentially game-changing tool, it also shows that a once-doubted machine-learning technique used to create the test can help develop real-world clinical solutions to intractable medical challenges. This interview with Khatri has been edited for length and clarity.

Why has sepsis been so difficult for physicians to diagnose?

Until recently, diagnostic tools have focused on finding the pathogen that caused the sepsis. But the immune system evolved expressly to keep pathogens out of the bloodstream because they're so deadly there, making it extremely difficult to find a pathogen in blood samples.

An additional complication is that sepsis can be either an overzealous body-wide immune system reaction to infection, or it can be the opposite: an insufficient response. For an older, obese male with immunosuppression, for example, sepsis is more likely to be caused by the body's inability to respond sufficiently to the infection.

People tend to think about sepsis as a binary problem, that the diagnostics are either you have sepsis or you don’t. But when you look at clinical practice, that’s not the right question. The real clinical question falls along two axes: 1) Is there infection or not? And when there is, is it bacterial or a viral? and 2) How severe a patient’s infection is likely to become, whatever its underlying cause?

What was the key insight or innovation that made the new test possible?

Researchers had already tried using transcriptomics – the study of copies of RNA instructions that tell cells what to do that are produced by the genome under specific conditions, providing insights into gene expression and regulation – to find gene activation profiles for infection. But finding a signal that generalizes to different populations under different conditions proved extremely difficult. That was because those studies were conducted on narrow populations – for example, children in hospital emergency departments in Kansas. But that same activation signature wouldn't work for, say, older adults in an intensive care unit in Haiti.

We knew the diagnostic would have to work in heterogeneous, real-world clinical environments. But enrolling a broad patient population to study was logistically daunting, in addition to being prohibitively expensive and time consuming. My “aha!” moment was realizing that we could use data that already exist in the public domain. Since everybody is sampling different patient populations for their studies, collectively, this data would represent the real-world heterogeneity we needed: different ages, races, sexes, ethnicities, eating habits, comorbidities, pathogen strains, and so on. Everyone thought of that heterogeneity as a curse, but we saw it as a blessing. We used methods I had developed to leverage the heterogeneity of that data to find the robust immune signatures that generalize to broad patient populations under all different conditions to identify the presence, type, and severity of an infection. They form the foundation of the sepsis test.

The statistical method we applied to all these divergent medical data sets is almost 50 years old. Our main innovation was broadening assumptions about when that method could be effectively applied. Originally, people suggested only studies that were approximately the same should be combined. But we combined studies that were as different from each other as possible to find signals that persisted through all the heterogeneity. We got unprecedented heterogeneity; we didn’t need to recruit any patients; and we used free, preexisting, public data.

That sounds like a smart solution to a seemingly intractable problem.

Well, not everybody liked the idea initially. In 2016, The New England Journal of Medicine ran an editorial calling those who use public data "research parasites," who don’t quite understand the original datasets. But we focused on leveraging heterogeneity that’s freely available. Anyway, we’ve gone from deep skepticism to FDA clearance!

How did you turn that insight, plus your machine-learning methods, into an actual test platform?

At that time, there was no platform that could measure tens of genes at the point of care rapidly enough that doctors wouldn’t just go ahead and start administering antibiotics as usual.

We decided to start a company and build the platform ourselves. Tim Sweeney was a general surgery resident at Stanford and a postdoc in my lab. In 2017, he quit all that he had going here to start a company and build the platform. The rest is history!

Tell me more about making the test.

After compiling as much public data as we could find from different countries, races, and ethnicities – age groups; genetic backgrounds; and different pathogen strains – we asked, “What is the set of genes that is always changing in the same direction if you have bacterial infection, or if you have viral infection, or you are likely to have a severe outcome from sepsis?” The machine learning methods that we developed found 29 genes that always change – under all those different conditions and in all those different populations – in the same directions.

The test provides three scores to answer three questions: Is there an infection? If yes, is the pathogen a bacterium or a virus? And, how severe is the infection? Within 30 minutes, an emergency department physician knows to treat for bacteria or for a virus, and whether to admit the patient or send them home.

How accurate is the test?

Each of the three scores is expressed as very low, low, moderate, high, or very high. When you look at the very high or the very low, in each of the three scores, our sensitivities and specificities are greater than 95%.

Also very important from a clinical point of view, we can reduce the number of false negatives in identifying severely ill patients. A false negative would be if a doctor examines someone in the emergency department and decides they do not need ICU-level care. If that patient did end up in the ICU, or if they died, that would be a false negative. Twenty-six people out of 1,200 who enrolled in the study for the FDA clearance did not receive ICU-level care, but died later. Of those 26, our test would have identified 18 as likely to be sick enough to admit to the ICU. In other words, where the current clinical practice has about a 2% false negative rate, ours is just 0.5%.

Are you still trying to improve those numbers?

Inflammatix is continuing to improve the test. In my lab, we have moved on to the next challenge – identifying the correct immune-modulatory treatment. Thanks to the TriVerity test, you’ll be able to be quickly and accurately diagnosed. But if the test says you have sepsis, we still don’t necessarily know what treatment is optimal.

This FDA clearance surely moves the sepsis test much closer to the clinic. But is it significant for other reasons?

Thirteen years ago, when I became an assistant professor, I realized that having access to real-world heterogeneity represented in publicly available data would enable us to develop new kinds of diagnostics that are generalizable to a broad patient population. Like I said, in those days we faced a lot of skepticism. We have gone from that skepticism to creating a groundbreaking test that has received FDA clearance. This is a proof of our concept and, we believe, the beginning of the end of the skepticism that analysis of public data can lead to important FDA-cleared, point-of-care diagnostics.