The statistics are troubling: Almost 49 million patients worldwide develop sepsis annually and about 11 million die from the complication. While there are several protocols for detecting it, these often fall short, which is why data scientists and developers are trying to create AI-driven algorithms to predict the condition early on. In previous blogs, we talked about the shortcoming of Epic’s sepsis algorithm, but since then several researchers have developed more promising digital tools in this domain.
Investigators at Johns Hopkins University have developed a machine learning system for sepsis called Targeted Real-Time Early Warning System (TREWS) and have used it in five hospitals over two years. When it was used to analyze over 9,800 retrospectively confirmed sepsis cases, it identified 82 percent of patients with the complication early on. Henry et al explain: “Adjusting for patient presentation and severity, patients with sepsis whose alert was confirmed by a provider within 3 h had a 1.85-h … reduction in median time to first antibiotic order compared to patients with sepsis whose alert was either dismissed, confirmed more than 3 h after the alert, or never addressed in the system.”
Of course, retrospective studies have weaknesses; they often fail to take into account confounding variables that might have skewed the findings. Fortunately, the Johns Hopkins team has also conducted a prospective multi-site study that looked at the value of the TREWS tool. That trial evaluated more than 6,800 patients who had been identified before antibiotic therapy was initiated.
They reported: “Adjusting for patient presentation and severity, patients in this group whose alert was confirmed by a provider within 3 h of the alert had a reduced in-hospital mortality rate… when compared with patients who alert was not confirmed by a provider within 3 h.”
The University of California at San Diego (UCSD) has also developed an AI tool that addresses sepsis. Their system, called COMPOSER, was used to predict the complication during a before-and-after prospective study performed at two EDs. The investigators evaluated over 6,000 adult patients with sepsis, relying on an advisory managed by their nurses. They found COMPOSER “was significantly associated with a 1.9 percent absolute reduction (17 percent relative decrease) in in-hospital sepsis mortality….”
Efforts to reduce the burden of sepsis using AI has not been limited to major US academic centers, however. Kim Huat Goh and associates, with Nanyang Technological University, Singapore, have developed an algorithm they call SERA, which incorporated structured data and unstructured clinical notes to predict and diagnose sepsis. They were able to predict the onset of the complication 12 hours ahead of time with an AUC of 0.94, sensitivity 0.87, and specificity of 0.87. Their data suggests that, compared to unaided physician predictions, the algorithm has the potential to “increase the early detection of sepsis by up to 32 percent and reduce false positives by up to 17 percent.”
A “different approach”
Other technologists are taking a somewhat different approach. Realizing the heterogenous nature of sepsis, they are developing algorithms that look for subtypes of patients who present differently from one another. Using a method called sequential organ failure assessment (SOFA), they create a score that evaluates several organ systems, including respiration, coagulation, liver, cardiovascular, central nervous system, and renal system functioning.
One group has isolated four phenotypic subgroups, each having distinct characteristics and outcomes. For example: “Patients with the β phenotype exhibit a higher prevalence of chronic illnesses and renal impairments, while those with the γ phenotype experience a greater incidence of inflammation and pulmonary dysfunction.”
Similarly, Luminare, Inc, which recently graduated from the Mayo Clinic Platform_Accelerate program, has detected seven phenotypes. Sarma Velamuri, MD, Luminare’s CEO, points out that one of the shortcomings of many AI systems is they use surrogate markers of the condition and generate binary outcomes: the disease is present or absent. For example, when training an AI model for sepsis detection, DRG codes and ICD-10 codes for sepsis are utilized or natural language processing is used to extract free text from unstructured data.
This approach results in low specificity and alert fatigue because the target being defined is binary. This methodology causes further problems because in clinical practice what is often deemed as sepsis from the AI model is not agreed upon at the bedside by providers. A better way to set the target is to phenotype patients into more granular categories, says Velamuri. That helps define the target better and helps with feature extraction. The method also enables you to fine-tune the AI model to look for adjacent clinical problems like early patient deterioration.
Luminare is also different from other sepsis evaluation systems because it does not directly predict which patient requires antibiotics or other interventions. Instead, it is a platform that provides staff nurses with a notification that educates them in-workflow on their patient, states which phenotype they are likely to fall into, and offers options for ordering the sepsis bundle inside the EHR. The nurse then decides how to use that information and can contact the appropriate physician through the platform — thus the likelihood of alert fatigue is diminished.
AI is making significant inroads in sepsis management. While these digital tools are far from perfect, they are moving us in the right direction.
This piece was written by John Halamka, MD, President, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform. To view their blog, click here.
Share Your Thoughts
You must be logged in to post a comment.