Background: Timely identification and treatment are integral to sepsis management in emergency department (ED) and inpatient settings. Increasingly, automated alerts embedded in electronic health records (EHRs) are used to expedite sepsis detection but may lead to false alarms and increased alarm fatigue. Additionally, recent data have demonstrated that the EPIC Sepsis Model—a proprietary sepsis alert widely implemented, often first fires after physician recognition of sepsis, thereby providing poor sensitivity of sepsis in comparison with clinical practice (Wong et al, JAMA IM, 2021). Our objective was to determine whether our institution’s home-grown EPIC EHR sepsis alert (Narayan et al, AJEM 2015) assisted providers in identifying sepsis based on timing of antibiotic ordering.
Methods: We included all adults who presented to a single academic ED between January 1, 2019 and June 30, 2021 with suspected sepsis, defined as having blood cultures ordered and intravenous antibiotics received within 48 hours of ED presentation. Demographic, laboratory, vital sign, clinical outcome, and medication data were obtained from the EHR along with date and clock timestamps. All EHR-based sepsis alerts occurring within 48 hours of ED presentation were extracted from the EHR along with their timestamp and the location in which the alert was triggered, ED versus inpatient. Briefly, our institution’s sepsis alert is an automated, real-time, algorithm-based sepsis alert that incorporates charted vital signs, laboratory values, and clinical assessments collected during the patient’s hospitalization. We calculated the time between documentation of the first sepsis alert and the first order for antibiotics.
Results: There were 9,990 ED patients that met inclusion criteria and 7,084 (70.9%) that had at least one EHR sepsis alert documented. Among those with an alert, most patients (78%, n=5,487) first had the sepsis alert documented and then had antibiotics ordered (“sepsis alert first”) (Figure 1). There were several clinical differences between patients who had the sepsis alert first versus those who had the sepsis alert after antibiotics were ordered (Table 1). Patients who had the sepsis alert first were more likely to be emergent on ED presentation (p< 0.0001), more likely to be triaged to critical care or transitional care at hospital admission (p=0.003), were more likely to have their alert triggered in the ED versus the inpatient setting (p< 0.0001), but had antibiotics ordered later (median 2.4 v. 1.4 hours, p< 0.0001), and administered later (median 3.3 v. 2.4 hours, p< 0.0001). In addition, patients with the sepsis alert first were more likely to die (8.8% v. 7.1%, p=0.032), even after adjustment for female gender, limited English proficiency, Elixhauser mortality score, ED acuity on presentation, admission location of care, time to antibiotic receipt, SOFA ≥ 2 within 48 hours, location of alert (ED versus inpatient), and ED census on the date of ED presentation (OR 2.3, CI 1.5-3.4). Patients who met inclusion criteria but did not have a documented sepsis alert were less acutely ill and had better outcomes than those who did have an alert documented.
Conclusions: Our institution’s home-grown EHR-based sepsis alert was successful at quickly identifying patients who were more acutely ill and more likely to die, particularly in the ED setting. However, the sepsis alert did not appear to have expedited antibiotic administration and its firing was not associated with improved survival among those with suspected sepsis.
