Year: 2021
Type: conference-abstract
Abstract: Introduction: Early sepsis detection may prompt earlier interventions and improved outcomes. Many hospitals employ proprietary, electronic health record (EHR)-based algorithms (e.g., the Epic Sepsis Prediction Model [ESPM]) to identify patients with sepsis at the time of, or before, sepsis onset. However, these models have not been consistently evaluated in real-world settings. We compared the ESPM's performance to that of two other common models in a multicenter cohort of hospitalized patients. Methods: We conducted a retrospective cohort study of adults 18 hospitalized between March 1, 2020, and June 30, 2020 at 11 BJC HealthCare hospitals. ESPM scores (probability [0-100] of sepsis during hospitalization based on real-time EHR data) were automatically calculated every 20 minutes for all inpatients but were not visible to clinical teams. We compared the ESPM, the systemic inflammatory response syndrome (SIRS), and the quick sepsis-related organ failure assessment (qSOFA) as predictors of three outcomes: suspicion of infection ( 1 culture from any site plus intravenous antibiotics), sepsis onset (suspicion of infection plus concurrent organ dysfunction), and hospital death. We estimated model discrimination (area under receiver-operator curve, AUROC), and, for multiple thresholds, positive and negative predictive values (PPV, NPV) for each outcome. Results: We evaluated 25,206 hospitalizations from 20,157 patients. Suspected infection occurred in 2,546 admissions. Of these, 1,861 involved sepsis and 349 resulted in hospital death. All models had poor discrimination for suspected infection (SIRS AUROC 0.54 [95% confidence interval (CI) 0.53-0.55]; qSOFA 0.62 [0.61-0.63]; ESPM 0.64 [0.63-0.65] and sepsis (SIRS AUROC 0.51 [0.50-0.53]; qSOFA 0.54 [0.52-0.55]; ESPM 0.60 [0.58-0.61], with the ESPM showing marginally better discrimination than qSOFA. Each model had fair discrimination for predicting hospital death, with the ESPM (AUROC 0.77 [0.75-0.80]) and qSOFA (AUROC 0.79 [0.77-0.82]) significantly outperforming SIRS (AUROC 0.72 [0.70-0.75], p < 0.05). Regardless of ESPM threshold, PPV was < 15% for all outcomes (Figure). Conclusions: The ESPM had poor discrimination for predicting the clinical onset of sepsis. Although ESPM discrimination for predicting hospital mortality was fair, it was not statistically different from that of qSOFA. The PPV for all models was low. Given the poor discrimination and PPV of the ESPM, the clinical utility of this model in its current form is low.
Authors: P.G. Lyons, B. Ramsey, M. Simkins, T.M. Maddox
Institutions: Washington University in St. Louis, BJC HealthCare
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W4244365944