For years, the approach to sepsis in acute care settings has been the same: early warning alerts and standardized care protocols.
Unfortunately, it hasn’t moved the needle. Sepsis still affects 2.5 million U.S. hospital patients every year, kills over 300,000, and costs the healthcare system $52 billion (AHRQ, 2021).
But what if the problem isn’t that we’re identifying sepsis too late? What if the problem lies in the inherent heterogeneity of sepsis? What if we looked further upstream at the underlying biology that drives each patient’s risk of sepsis?
That shift in thinking led to something unprecedented: the first ever FDA-authorized AI diagnostic tool for sepsis.
Why Biology Matters
Sepsis isn’t one disease. Rather, it’s a dysregulated immune response that looks completely different from patient to patient. Some become hyperinflamed. Others slip into immune suppression. Some patients cycle through both in unpredictable ways.
“In the past, there have been a lot of failed attempts with therapies that have gone broadly after patients with sepsis who have different kinds of underlying biology,” says Nathan Shapiro, professor of Emergency Medicine at Harvard Medical School. “The idea going forward is to uncover the underlying biology and then add a therapeutic specifically targeted to the problem for an individual patient.”
Traditional scoring systems like SOFA, NEWS, and qSOFA rely on clinical thresholds that can indicate many conditions, not just sepsis. Single biomarkers offer only a narrow snapshot of what’s happening. Early warning systems issue alerts for anyone who could have sepsis. The result is noise and uncertainty on the best course of treatment for each patient.
A New Solution, Rooted in Research
Over the course of a decade, researchers at Prenosis built a biobank containing over 130,000 blood samples from more than 30,000 hospitalized patients. They studied dozens of biomarkers not commonly associated with sepsis, looking for patterns that revealed true risk.
The result was Sepsis ImmunoScore: an algorithm analyzing 22 parameters including procalcitonin, C-reactive protein, white blood cell counts, and vital signs.
The team pursued FDA authorization from the start, conducting multicenter validation across six hospitals. Their results, published in NEJM AI, showed diagnostic accuracy with AUCs ranging from 0.80 to 0.85. More importantly, the tool accurately predicts who will progress to sepsis, ICU admission, mechanical ventilation, or death within 24 hours.
“It’s hard for the human brain to integrate 22 separate factors in tandem, but machine learning helps give [clinicians] that 3-D clinical picture, combining parameters of patient biology to reveal the risk level of a patient so the clinician can enhance their decision-making,” says Akhil Bhargava, lead author on the NEJM AI paper. “It’s a tool to help providers who want a diagnostic test to know the relative risk of sepsis for a patient.”
Practical Impact for Acute Care
The Sepsis ImmunoScore AI diagnostic tool can work in tandem with early warning alert systems, and within already existing clinical workflows. And the applications extend beyond augmenting clinical decision-making. Early data suggests the tool could reduce antibiotic overprescribing in non-infected patients, can help hospitals improve SEP-1 bundle compliance, and assist clinical teams with appropriate allocation of ICU resources.
The biobank research is now focused on patient stratification based on underlying biology—work that could finally enable the targeted sepsis therapies that have eluded researchers for decades. “The major limitation in the field is that sepsis is a human construct that actually encompasses a wide variety of pathophysiological processes happening in patients,” says Greg Watson, PhD, senior machine learning scientist at Prenosis and statistical advisor for the NEJM AI study. “The underlying biological heterogeneity has thwarted attempts to develop effective treatments for patients because we’re filing them all under the umbrella of sepsis. [We are] working really hard to amass and harness deeply informative biological and clinical data so we can learn from it and push the acute care field forward.”
Get the Full Story
The journey from concept to FDA authorization involved more than just algorithm development. It required rethinking what sepsis diagnosis should look like.
Read the complete story in the white paper: Know Thine Enemy: Why Biology-Based Sepsis Diagnosis Equals Real Progress at prenosis.com.
