A clinical prediction model to identify patients at high risk of death in the emergency department

Abstract

Purpose: Rapid assessment and intervention is important for the prognosis of acutely ill patients admitted to the emergency department (ED). The aim of this study was to prospectively develop and validate a model predicting the risk of in-hospital death based on all available information available at the time of ED admission and to compare its discriminative performance with a non-systematic risk estimate by the triaging first health-care provider.

Methods: Prospective cohort analysis based on a multivariable logistic regression for the probability of death.

Results: A total of 8,607 consecutive admissions of 7,680 patients admitted to the ED of a tertiary care hospital were analysed. Most frequent APACHE II diagnostic categories at the time of admission were neurological (2,052, 24%), trauma (1,522, 18%), infection categories [1,328, 15%; including sepsis (357, 4.1%), severe sepsis (249, 2.9%), septic shock (27, 0.3%)], cardiovascular (1,022, 12%), gastrointestinal (848, 10%) and respiratory (449, 5%). The predictors of the final model were age, prolonged capillary refill time, blood pressure, mechanical ventilation, oxygen saturation index, Glasgow coma score and APACHE II diagnostic category. The model showed good discriminative ability, with an area under the receiver operating characteristic curve of 0.92 and good internal validity. The model performed significantly better than non-systematic triaging of the patient.

Conclusions: The use of the prediction model can facilitate the identification of ED patients with higher mortality risk. The model performs better than a non-systematic assessment and may facilitate more rapid identification and commencement of treatment of patients at risk of an unfavourable outcome.

Fig. 1

Calibration plot showing a all deciles and b the lower nine deciles of predictions. Predicted probabilities using the model’s coefficients are aggregated to deciles of patients. The mean observed death rate in each decile is the percentage (and 95 % confidence intervals) of observed death from all observations in this decile, and are marked by black triangle and lines. The dashed line indicates the optimal 1:1 fit. The dotted line represents the locally weighted scatterplot smoothing (LOWESS) smoother of the predictions. The calibration slope was 0.95, indicating good calibration

Fig. 2

ROC curve of the final model. Area under the curve (AUROC) was 0.922 (95 % range 0.916–0.927), indicating good internal validity

Fig. 3

Comparison of prediction of risk of death of the nurse risk estimate model and the developed prediction model in deciles of patients according to the calibration plot. In the 8 deciles of patients with lower mortality risk the nurse risk estimate consistently predicted a higher risk of death than the developed model with considerable higher variance in prediction. In the 20 % of patients with the highest risk of death, higher mortality risk was predicted by the model than by the nurse risk estimate. N nurse risk estimate model, M prediction model