CLN Article
In sepsis management, time is of the essence. Every hour that passes before a patient is identified and started on treatment puts them at risk for organ failure or death. Knowing this, some clinicians have been trying new methods to increase earlier sepsis detection and management by developing algorithms that employ artificial intelligence (AI) to evaluate laboratory values and other data to alert providers to the presence of sepsis, which affects some 50 million people globally. Although the performance of these algorithms varies so far, this overall trend presents a perfect opportunity for laboratory medicine professionals to showcase their knowledge.
AI is just one of several solutions emerging for sepsis detection, thanks in part to biomarkers for the condition that can be measured in the blood, such as circular RNAs, protein C, and prokineticin 2 (Int J Mol Sci 2024; doi.org/10.3390/ijms25169010), said Damien Gruson, PhD, head of the department of clinical biochemistry at Cliniques Universitaires Saint-Luc in Brussels, Belgium, and chair of the Division on Emerging Technologies of the International Federation of Clinical Chemistry and Laboratory Medicine. Other emerging solutions include point-of-care testing and omics options such as genomics and high-throughput sequencing.
“All together, these tools should help clinical teams to better diagnose and manage the patient,” Gruson said. “It’s not a substitution of the healthcare workforce by the technology. It’s the amplification of human intelligence by artificial intelligence.”
For example, a multidisciplinary team at the University of California San Diego (UCSD) School of Medicine developed COMPOSER (COnformal Multidimensional Prediction Of SEpsis Risk), a deep learning model for early sepsis prediction. Once a patient checks into the emergency department, the algorithm starts monitoring 150 types of datapoints, including lab results, vital signs, home medications, comorbidities, and demographics to look for possible signs of sepsis. If the program suspects a person is septic, or heading that way, it sends an alert through the electronic health record (EHR) to the nursing staff, who can then review these values with the physician to confirm and determine treatment.
In a recent study (NPJ Digit Med 2024; doi: 10.1038/s41746-024-01234-1), UCSD clinicians reviewed records from 6,217 adult septic patients presenting to two of its hospital emergency departments, before and after the COMPOSER tool was deployed. The use of the deep learning tool was associated with a 17% relative decrease in in-hospital sepsis mortality and a 10% relative increase in adhering to an evidence-based management guideline known as the sepsis bundle.
They also found less organ injury 72 hours from the time of sepsis detection. Once the tool was activated, it generated an average of 235 alerts per month, corresponding to 1.65 alerts per nurse per month. In more than half of cases, a nurse responded that they would notify the physician immediately.
The idea behind the program was to help identify sepsis in cases where it might not be obvious, said Gabriel Wardi, MD, MPH, an associate professor of emergency medicine and chief of the division of critical care at UCSD, who helped develop the program.
“If someone comes in with a very low blood pressure, high fever, high heart rate — that will be picked up immediately,” said Wardi. “Where the tool adds value is where there’s diagnostic uncertainty with these patients.” People with vague symptoms often have significant delays with their treatment, he said.
Researchers at other academic medical centers including Duke, the University of Pennsylvania, and Johns Hopkins also have designed AI-enabled algorithms to detect sepsis, with promising results. But not all tools have fared as well. The EHR firm Epic offers its own sepsis detection model, but when investigators at the University of Michigan went to validate it in some of their hospital units, they found the tool correctly sorted patients on their risk of sepsis only 63% of the time, not 76%–83% of the time as reported by the company. The tool also sent out too many alerts regarding patients who turned out not to have sepsis (JAMA Intern Med 2021; doi:10.1001/jamainternmed.2021.2626). The researchers determined that the model was trained on data defining the onset of sepsis as the time a clinician intervened, not necessarily when it started, which skewed results.
Clinicians in St. Louis determined the Epic model wasn’t a fit for them for similar reasons, said Ronald Jackups Jr., MD, PhD, chief medical information officer for laboratories at BJC Healthcare in St. Louis and professor of pathology and immunology at Washington University School of Medicine. As such, he and his colleagues are not using any AI tools for sepsis detection.
It is a good reminder that all such tools need to be validated by multidisciplinary groups including frontline clinicians, data scientists, information technology specialists, and others before adopting them, especially if they’re being adopted across settings. Laboratorians should be key members of such teams, said Jackups, president of the Association of Pathology Informatics (API). Laboratory leaders also should take an active role in the development of such programs and get involved with hospital committees and initiatives surrounding identification and treatment of sepsis.
“Those groups frequently use lab test results to drive these decisions but may not directly reach out to laboratory leaders to include them as subject matter experts,” he said. “It is our job to reach out and make sure we’re included in those discussions. That gives us the opportunity to provide our expertise on the utility of these tests: when they should be ordered, how they should be interpreted, and whether alerts are valuable in this process.”
API partnered with Project Santa Fe Foundation on a new Diagnostic Medicine Consortium aimed at maximizing the predictive value of information generated by diagnostics.
“One of the major goals for the future of laboratory medicine is to get out of the laboratory and start assisting the front-facing clinical services with not just understanding lab tests, but applying them to improve patient care,” he said. “This is exactly one of those kinds of initiatives where we can show that.”
To help, clinical laboratories can work to standardize their data as needed to allow comparisons among hospitals or within different areas of the hospital, Gruson said. They also could incorporate diagnostic testing for biomarkers and host response to infection, said Wardi. In the case of expensive tests, AI programs potentially could help clinicians and laboratories prioritize which patients to run tests on, “which allows us to both manage costs and improve care,” said Suchi Saria, MSc, PhD, the John C. Malone Associate Professor of computer science at the Johns Hopkins Whiting School of Engineering in Baltimore, and associate professor of statistics and health policy at Hopkins’ Bloomberg School of Public Health. Saria also is the founder of Bayesian Health, a clinical AI platform company spun out of her university research on the TREWS (Targeted Real-Time Early Warning System) sepsis detection tool, which has demonstrated an 18.7% reduction in sepsis mortality (Nature Medicine 2022; doi.org/10.1038/s41591-022-01894-0).
AI models for sepsis and other conditions could become commonplace within the next 5 to 10 years, Jackups said. But there are several hurdles to more widespread use. One is trust in output from alerts where users don’t understand the complex algorithm driving the program, known as the “black box.” The Food and Drug Administration is one of several groups that have attempted to establish guidelines.
“The most important guideline is that the alert shouldn’t be making the decision — the alert should be providing enough information and knowledge so that the ordering provider can make an appropriate decision,” Jackups said. That includes presenting as much useful information on why an alert occurred and what the next steps could be, without dictating those steps. “Those kinds of alerts are much more likely to be understood and followed by clinicians, as opposed to ‘black box’ alerts that simply give a directive without fully explaining why the alert is even occurring.”
Another hurdle is how to build models to be early and accurate in their data while also driving adoption, Saria said. Ideally, programs need to be integrated into the EHR and workflows and work across settings from critical care to the emergency department and beyond. Saria’s approach has resulted in a nearly 90% adoption rate among providers (Nature Medicine 2022; doi.org/10.1038/s41591-022-01895-z).
“You can have a perfectly validated tool that has really high sensitivity and positive predictive value and then completely fail to change practice when implemented,” Jackups added. “It is critical for someone with an understanding of change management to be able to observe how providers use these kinds of alerts and react to them. That culture is institution-dependent.”
Wardi found a generational divide: Younger colleagues were excited about using the COMPOSER tool, while some of the senior physicians essentially said, “thanks but no thanks.” He sat down with some of them individually to encourage them to think of the tool as a second set of eyes in a busy emergency department, and if an alert went off to at least take another look at their patient. “That actually worked quite well,” he said.
Overall, said Wardi, “Where I see this going is to have an emphasis on multimodal data, which isn’t just getting data from the EHR but rather bringing in the clinical notes, and novel tests to assess host response to infection and develop a whole new kind of domain of data that will hopefully — at least from what I’ve seen based on our preliminary work — really improve how these models perform to become much more useful for clinicians to and give them a very high predictive value.”
The technology also could be used cross-purpose for multiple other conditions, from earlier identification of pressure ulcers to helping determine which patients would benefit from palliative care, Saria said.
“I see this huge opportunity for what I call intelligent care augmentation,” she said. “We’re taking AI and a smart, real-time platform that augment the capabilities of the care team, which in turn helps them parse data to identify high-quality, validated, robust clinical signals that are more proactive than reactive, and enables smart workflows. It enables them to give the right targeted treatments and provide better, faster, and smarter care — while also saving them a huge amount of time.”
Karen Blum is a freelance medical and science writer in Owings Mills, Maryland. +Email: [email protected]