PURPOSE: The cornerstone of sepsis treatment is early goal directed therapy (EGDT), however, delayed recognition is an important barrier to early treatment. We developed the capacity to perform continuous, automated, real-time surveillance to identify patients at risk for sepsis using the electronic medical record.
METHODS: Between July 6 and September 30, 2009, patients >18 years of age who were admitted to a non-ICU location without a Do-Not-Resuscitate order were monitored to identify abnormalities suggestive of sepsis. An alert was generated, but not clinically disseminated, if ≥2 SIRS criteria were present in conjunction with ≥1 abnormality suggestive of organ dysfunction. Using only a patient’s first admission and sepsis alert, multivariable logistic regression was used to predict in-hospital mortality while controlling for various demographic characteristics. Sensitivity and specificity were calculated using administrative ICD-9 discharge diagnoses (Dombrovskiy definition).
RESULTS: During the study period 7247 patients met the inclusion criteria, 1559 (21.5%) had ≥1 alert, 239 (3.2%) had a diagnosis of sepsis (2.0%), 145 had a diagnosis of severe sepsis, and 239 (3.2%) died. One-third of all in-hospital deaths occurred in patients with an administrative diagnosis of sepsis. In-hospital mortality for alert and non-alert patients was 7.3% and 2.2%, respectively (p<.0001). The multivariable model demonstrated that a sepsis alert (OR 2.8, CI95% 2.2, 3.7), age > 45 (OR 2.3; CI95% 1.5, 3.3), age >65 (OR 3.9, CI95% 2.6, 5.7), admission from the ED (OR 5.2, CI95% 3.5, 7.9), and admission from another institution (OR 9.7, CI95% 6.3, 15.0) independently predicted in-hospital mortality. The sensitivity of the alert for the administrative diagnosis of sepsis and severe sepsis was 71.6% and 72.3%, respectively. The specificity of the alert for sepsis and severe sepsis was 80.1% and 79.5%, respectively. The PPV for a sepsis and severe sepsis diagnosis was 10.6% and 6.9%, respectively.
CONCLUSIONS: The sepsis alert independently predicted in-hospital mortality.
CLINICAL IMPLICATIONS: While its sensitivity is reasonably high, its relatively low specificity and the low sepsis prevalence yield a considerable false positive rate.
DISCLOSURE: Jorge Alsip: Employee: Employee of Cerner Corporation that markets electronic medical records to health care institutions. I am a consultant to UAB on this project and participated in its design and implementation.
The following authors have nothing to disclose: Joe Gerald, Joan Hicks, Michael Waldrum, Nancy Dunlap
This abstract describes the outcomes of a clinical decision rule that is embedded in a proprietary EMR system (Cerner Millinium). The decision rule is not specifically a product of Cerner but rather was developed in collaboration with them. This decision rule has not been "approved" for use but is being used by other health care institutions in the United States in a form that is very similar, but not exactly the same as the rule described in this abstract.