PREDICTing Mortality in the Emergency Department: External Validation and Derivation of a Clinical Prediction Tool. (29th May 2017)

• Record Type: Journal Article
• Authors: Moman, Rajat N.; Loprinzi Brauer, Caitlin E.; Kelsey, Katherine M.; Havyer, Rachel D.; Lohse, Christine M.; Bellolio, M. Fernanda
• Editors: Alter, Harrison J.
• Is Part Of: Academic emergency medicine. Volume 24:Number 7(2017)
• Page Start: 822
• Page End: 831
• Publication Date: 2017-05-29
• Subjects: Emergency medicine -- Periodicals; 616.02505
• DOI: 10.1111/acem.13197
• Languages: English
• ISSNs: 1069-6563

Abstract

Abstract: Background: The Choosing Wisely campaign has called for better engagement of palliative and hospice care services for patients in the emergency department (ED). PREDICT is a clinical prediction tool that was derived in an Australian ED cohort. It assesses a patient's risk of mortality at 1 year to select those who would benefit from advanced care planning. Such goals‐of‐care discussion can improve patients' ability to communicate what they want out of their healthcare and, in cases of end of life, potentially reduce the number of futile interventions. Using a cutoff of 13 points, PREDICT had a reported 95.3% specificity and 53.9% sensitivity for 1‐year mortality. We externally validated PREDICT and derived a simpler modified PREDICT tool to systematically identify high‐risk patients eligible for goals‐of‐care discussions and palliative care consultation in the ED. Methods: This was an observational cohort study of a random sample of 927 patients aged 55+ seen in the ED in 2014. We identified advance healthcare directives (AHDs) on file. We summarized diagnostic accuracy of the clinical tool to predict 1‐year mortality using sensitivity, specificity, and area under the curve (AUC). We refined PREDICT using multivariable modeling. We followed reporting guidelines including STrengthening the Reporting of OBservational studies in Epidemiology (STROBE) for cohort studies and Standards for Reporting of Diagnostic Accuracy (STARD). Results: A total of 927 patients were included: 55.0% were male, 63 (7.0%) were nursing home residents, 389 (42.0%) patients had an AHD in their medical record at the time of ED visit, and 245 (26.4%) were deceased at 1 year. Of the 780 patients with PREDICT scores < 13, a total of 164 (21.0%; 95% confidence interval [CI] = 18.3–24.1) were deceased at 1 year, and of the 147 patients with PREDICT scores ≥ 13, a total of 81 (55.1%; 95% CI = 46.7–63.2) were deceased at 1 year. The AUC of the PREDICT score was 0.717 (95% CI = 0.680–0.754), sensitivity was 33.1% (95% CI = 27.3–39.4), and specificity was 90.3% (95% CI = 87.8–92.4) to predict 1‐year mortality. The modified PREDICT tool resulted in an AUC of 0.709 (95% CI = 0.671–0.747). We decided to select this model as the preferred model, as the variable of intensive care unit (ICU) admission with multiorgan failure can be difficult to assess in the ED and may delay advanced care planning. Reweighting the score did not improve fit or the AUC, so points assigned to each variable were not adjusted. Conclusion: PREDICT is an easy tool to administer to be able to identify patients who are at high risk of 1‐year mortality and who could benefit from AHDs, goals‐of‐care discussion, and when appropriate in the context of an end‐of‐life setting, palliative medicine consultation. External validation of PREDICT was successful in our population. We simplified PREDICT and derived a new tool, the modified PREDICT minus ICU tool, without significantly altering the sensitivity, specificity, and AUC for death at 1 year. The next steps include external validation of the newly derived rule and prospective implementation.