Health09 April, 2020

Sepsis, COVID and the growing need for AI-driven surveillance

As hospitals and health systems scramble to contain the COVID-19 pandemic, sepsis has emerged as an important consideration.

A recent study of a Chinese hospital found that 100 percent of patients who died there had sepsis, and there are reliable indications that sepsis burden will only increase in 2020.

Long before the current crisis, governing bodies, clinical experts and hospitals were well aware of the danger sepsis poses. As we highlighted in our recently published guide for sepsis coordinators, over the past two decades, increasingly sophisticated guidelines have emerged from the CDC, the Society for Critical Care Medicine and others that have helped hospitals recognize sepsis earlier and treat it more effectively. Clinicians should also follow these guidelines for sepsis in COVID-19 patients.

The guidelines have engendered progress, but even before COVID, studies indicated that sepsis still accounts for nearly 270,000 deaths each year in the United States – about 1 in 3 of all hospital deaths. (Worldwide, according to The Lancet’s 2020 Global Burden of Disease Study, sepsis accounted for almost 20 percent of all deaths in 2017.) When sepsis patients don’t die, they are susceptible to numerous clinical problems that have devastating effects, including gangrene, which too often leads to the loss of limbs, as well as respiratory failure, coma, and kidney failure, which can force patients onto extended periods of life support, mostly in ICUs that for the foreseeable future will be severely impacted.

It’s also important to add that only about half of sepsis patients get CMS-recommended evaluations and treatments within the first six hours. Commonly, clinical staff members fail to recognize sepsis right away and each hour delay in giving antibiotics may increase the mortality rate by about 8 percent.1 Furthermore, the corresponding financial costs in the US are $62 billion a year and the reimbursement rate for each sepsis incident typically leaves hospitals on the hook for somewhere between $6,500 and $12,000.

Taken together, these stark facts demonstrate the pressing need for more timely detection, diagnosis, and treatment. Today’s cost in human lives and hospital resources remains unacceptable.

What’s slowing progress?

Existing systems for automating sepsis detection tend to use SIRS (systemic inflammatory response syndrome) criteria—an extremely sensitive measure that identifies nearly all potential sepsis cases, but also generates a lot of false positives, resulting in alert fatigue. These systems look for key indicators of physiological stress, such as an accelerated heart or respiratory rate and while sepsis certainly causes those abnormalities, so does any significant form of stress. A recent study found that approximately 90 percent of hospitalized individuals have physiological signs of stress that meet SIRS criteria.2

The resulting number of false positives—or in the case of other measures like SOFA (sepsis-related organ failure assessment), the risk of missing sepsis cases or identifying them too late in the process—speaks to the need for a more sophisticated clinical surveillance tool, one that can rapidly identify nearly all potential cases, but only those that are likely to be sepsis. Such a tool must cull through and make sense of enormous data volumes and then deliver a timely alert to the point-of-care where a clinician has what he or she needs to diagnose and treat. The data comes from a variety of locations and in a number of forms; the most important data tends to be unstructured data buried within clinician’s notes, inside of a patient’s electronic health record.

This is precisely the type of challenge that artificial intelligence (AI)—in particular, natural language processing (NLP), a form of AI—is designed to solve. Done right, AI combs through various data types and normalizes it at a speed that human beings cannot replicate, thereby transforming it into discrete data that is accessible for analysis.

Yet to date, bringing AI/NLP to healthcare has been a challenge because of the difficulty of understanding the unique language of the clinical domain. Under intense time pressure, clinical notes are often full of parentheticals, incorrect grammar, clinical jargon and acronyms. Moreover, the precise clinical jargon and acronyms can often vary from setting to setting, from specialty to specialty and from clinician to clinician. It has taken some time to address this complexity, but the solution is now at hand.

How NLP delivers timely alerts to clinicians at the point of care

As noted above, clinical notes hold much of the essential information for diagnosis and treatment of sepsis, including differentiating the patient with sepsis from a patient with a condition that has some of the same characteristics.

Consider a recently hospitalized patient who has been on oxygen at home due to emphysema. Even when this patient is in her normal state at home, she may develop tachycardia and shortness of breath when she gets up to use the bathroom. Without understanding her baseline state, a screening tool may incorrectly interpret her symptoms, but a surveillance tool that can understand clinical notes can recognize that those symptoms are due to emphysema, not sepsis. Additionally, if this same patient develops a COPD exacerbation, an NLP-driven tool will know to expect persistent tachycardia, tachypnea and mild hypoxia. The tool can also anticipate the patient might receive albuterol and corticosteroids, which will further accelerate the heart rate and cause leukocytosis, respectively. In both of these scenarios, a strong AI surveillance tool that leverages NLP can prevent an unnecessary alert.

If a clinician has already determined that a patient does have sepsis and documented sepsis in his or her note, an NLP-driven tool can recognize that the provider knows the patient has sepsis and opt not to send an alert to minimize alert fatigue.

These are very sophisticated determinations, which is why technology experts alone cannot produce such a clinically tuned tool. In the case of Wolters Kluwer’s POC Advisor, a clinical surveillance solution that deploys NLP, clinicians were central to training the complex algorithms that effectively integrate the various types of data and deliver timely and accurate alerts to the point of care. According to a peer-reviewed study in the Journal of the American Medical Informatics Association, use of POC Advisor correlated with a reduction in sepsis mortality of 53 percent, and a reduction in related 30-day readmissions of 31 percent.3 NLP refinements since that study appear to have enhanced both the sensitivity and specificity of the tool with customers seeing approximately 99% sensitivity and 97% specificity for alerting

This is a particularly important development, because it has implications that go far beyond sepsis. For example, some patients with COVID rapidly develop acute respiratory distress syndrome or ARDS, which requires emergency care. Like sepsis, this is a life-threatening event, so if it is not detected expediently, the consequences can be dire. POC Advisor can now provide alerts to clinicians for COVID patients who need emergency clinical attention. We can also apply some of the same surveillance principles to detect C. diff infections, over-sedation, severe pneumonia and hemorrhage. The opportunities are endless.

But integrating such a tool effectively in health care’s complex environment demands more than the tool itself. Change management expertise is essential, especially to build trust among the clinicians who must rely on the tool. Two elements are particularly important.

First, demonstrate success. Choose a discrete clinical challenge and show how an AI-driven tool can help. Once clinicians see the reliability and efficiency in one setting, they are more willing to extend the capabilities to others. Second, identify your change agents and champions. They are the ones who create policies and order sets that support the surveillance results. They are the ones who educate staff. And they are the ones who analyze results, share them with clinicians and continually revisit and refine processes until they are working as smoothly as possible.

The coronavirus is exacting unprecedented demands on healthcare systems around the world that were already struggling to improve clinical outcomes and survive financially. Now these same systems are being stretched to the limit, scrambling for scarce resources and forcing providers, who are putting their lives on the line, to learn on the fly about how to detect and treat a once-in-a century threat, even as they continue to wrestle with established threats, like sepsis. In this context, an AI-driven surveillance tool can become a lifesaver.


Ferrer, Ricard MD, PhD1; Martin-Loeches, Ignacio MD, et al. “Empiric Antibiotic Treatment Reduces Mortality in Severe Sepsis and Septic Shock From the First Hour: Results From a Guideline-Based Performance Improvement Program,” Crit Care Med. 2014 Aug;42(8):1749-55.

Churpek, Matthew M., et al. "Incidence and prognostic value of the systemic inflammatory response syndrome and organ dysfunctions in ward patients." American journal of respiratory and critical care medicine 192.8 (2015): 958-964.

Manaktala, Sharad, et al. “Evaluating the impact of a computerized surveillance algorithm and decision support system on sepsis mortality.” JAMIA, 24(1), 2017, 88-95.

Back To Top