Ward death following ICU discharge - can it be predicted? Development and validation of a predictive model

Full text

Introduction

A significant number of patients discharged alive from the intensive care unit (ICU) die subsequently on the general wards; a problem compounded by early discharge due to pressure on ICU beds. We analysed data from the patients' last day in the ICU prior to discharge, during their first ICU admission, to derive a predictive model that would identify those patients who were at risk of dying on the ward.

Patients and methods

All ICU survivors discharged between 1st June 1990 to 31st December 1996 were included in the study. Of the 5475 survivors there were 200 (3.7%) ward deaths. Physiological and treatment data, collected prospectively from the Intensive Care Database (Riyadh Intensive Care Program, Medical Associated Software House Ltd, London, UK), for the patients' last day in the ICU were analysed to determine which variables had a significant influence on death following ICU discharge. Multiple logistic modelling was used to select which variables produced the `best' predictive model of ward death. The model was then validated using two independent data sets.

Results

A higher mean acute physiology score (APS) (10.1 vs 8.0), increasing age (65 vs 59 years), presence of chronic ill health (39% vs 22%), longer ICU stay (8 vs 3 days), and whether the patient had cardiac surgery, were selected for inclusion in the model. The APS, age, chronic ill health and length of ICU stay were positively associated with death whilst having had cardiac surgery was positively associated with survival. Using a 0.6 cutoff the predictive model demonstrated good discrimination (sensitivity 66%, specificity 88%, ROC curve 85.6%) and calibration (Χ² 7.5982; P=0.4737). External validation of the model produced a ROC curve of 86.8% (sensitivity 90%, specificity 68%) (data set 1) and 77.5% (sensitivity 75% specificity 69%) (data set 2).

Conclusion

Using patient data obtained at ICU discharge it is possible to develop a discharge triage model which identifies those patients who are at risk of dying on the ward following ICU discharge.