Early antibiotic therapy has been shown to improve outcomes in sepsis, but patients often present with vague symptoms, making it difficult to recognize the condition early. Treatment delays can occur as patients wait for their initial evaluation, while laboratory tests are processed, or because of other more urgent cases. “Despite previously established quality measures to improve outcomes, timely sepsis treatment remains a widespread challenge,” says Rashid Bashir, Ph.D. Hospitals have challenges to meet quality measures, and even after a sepsis diagnosis is made, it can take time to alert the entire healthcare team to prioritize treatment.

In published studies, researchers have described a variety of methods to screen for sepsis and respond earlier in the disease course by providing timely antibiotics. These interventions range from simple scoring systems to more complex multivariable algorithms. Although these tools shed light on which patients should be strongly considered for antibiotics, they do not consider the risks and benefits clinicians must weigh when deciding to initiate treatment. In addition, available tools are limited in that they cannot identify or predict the disease course in sepsis.

Testing Applied Machine Learning Using EMR Data

With the availability of large volumes of EMR data, machine-learning methods are being developed to identify patients with sepsis early. Dr. Bashir and colleagues conducted a study, published in Clinical & Translational Science, in which they applied machine learning using EMR data and 15 novel biomarkers that are uncommonly measured to identify patients with sepsis.

Specifically, the authors presented a more holistic analysis of a subset of a large dataset comprised of three plasma proteins—procalcitonin (PCT), interleukin-6 (IL-6), and C-reactive protein (CRP)—and routinely measured EMR parameters. PCT, IL-6, and CRP have shown strong predictive power for sepsis and sepsis-related outcomes. In total, the study team evaluated sample collections from 1,400 adult patients in EDs who were suspected to have sepsis.

Demonstrated Diagnostic Capabilities & Prognostic Power

The machine-learning model rapidly identified patients with sepsis and stratified them by severity at the time when a first clinical specimen was acquired. The model yielded a score with diagnostic capability as well as prognostic power concerning hospital length of stay (LOS), 30-day mortality, and 3-day inpatient readmissions in the entire testing cohort and in various subpopulations. The area under the receiver operating curve for diagnosing sepsis was 0.83.

Predicted risk scores for patients with septic shock were higher than scores were seen among those with sepsis but without shock. Predicted risk scores for patients with infection and organ dysfunction were higher than what was seen in those without either condition. The authors noted that stratification based on predicted scores of patients into low, medium, and high-risk groups showed significant differences in LOS, 30-day mortality, and 30-day inpatient readmission (Table).

Providing Guidance for Clinicians on Appropriate Treatment

The machine-learning model may effectively identify problematic patient groups with vague presentations who need prompt broad-spectrum antibiotics earlier in the disease course, according to Dr. Bashir. “The algorithm distinguishes septic patients from those with other non-descript features, meaning it is less likely to misclassify individuals who have uncomplicated infections or organ dysfunction due to a non-infectious process,” he says. Of note, PCT and IL-6 are among the most important features in the model. These biomarkers are useful in differentiating patients with organ dysfunction and infection.

A tool to identify patients with sepsis and those who will soon develop it may facilitate appropriate treatment, notes Dr. Bashir. “Our machine learning-based score used clinical data from EMRs and three biomarkers that are not routinely measured to improve diagnostic performance,” he says. “These data help differentiate patients with sepsis while reflecting the severity of illness.” He adds that the early identification tool can be used to guide clinicians and healthcare staff by facilitating the prioritization of broad-spectrum antibiotic administration for medium- or high-risk patients, which in turn, has the potential to improve clinical outcomes.