OBJECTIVE: To develop a prediction model of 28-day mortality in adult intensive care units using administrative data. DESIGN, SETTING AND PARTICIPANTS: We obtained data from 33 ICUs in Japan on all adult patients discharged from ICUs in 2007. Three predictive models were developed using (i) the five variables of the Critical Care Outcome Prediction Equation (COPE) model (age, unplanned admission, mechanical ventilation, hospital category and primary diagnosis) (the C model); (ii) 11 variables, including the COPE variables and six additional variables (sex, reason for ICU entry, time between hospital admission and ICU entry, use of fresh frozen plasma or a platelet preparation, dialysis, and use of pressors/vasoconstrictors (the P+ model); and (iii) ten of the 11 variables, excluding primary diagnosis (the P- model). Data for 6758 patients were stratified at the hospital level and randomly divided into test and validation datasets. Using the test dataset, five, 10 or nine variables were subjected to multiple logistic regression analysis (sex was excluded [P > 0.05]). MAIN OUTCOME MEASURE: Mortality at 28 days after the first ICU day. RESULTS: Areas under the Receiver Operating Characteristic curve (AUROCs) for the test dataset in the C, P+ and P- models were 0.84, 0.89 and 0.87, respectively. Predicted mortality for the validation dataset gave Hosmer-Lemeshow chi2 values of 12.91 (P = 0.12), 10.76 (P = 0.22) and 13.52 (P = 0.1), respectively, and AUROCs of 0.84, 0.89 and 0.90, respectively. CONCLUSIONS: Our P- model is robust and does not depend on disease identification. This is an advantage, as errors can arise in coding of primary diagnoses. Our model may facilitate mortality prediction based on administrative data collected on ICU patients.
OBJECTIVE: To develop a prediction model of 28-day mortality in adult intensive care units using administrative data. DESIGN, SETTING AND PARTICIPANTS: We obtained data from 33 ICUs in Japan on all adult patients discharged from ICUs in 2007. Three predictive models were developed using (i) the five variables of the Critical Care Outcome Prediction Equation (COPE) model (age, unplanned admission, mechanical ventilation, hospital category and primary diagnosis) (the C model); (ii) 11 variables, including the COPE variables and six additional variables (sex, reason for ICU entry, time between hospital admission and ICU entry, use of fresh frozen plasma or a platelet preparation, dialysis, and use of pressors/vasoconstrictors (the P+ model); and (iii) ten of the 11 variables, excluding primary diagnosis (the P- model). Data for 6758 patients were stratified at the hospital level and randomly divided into test and validation datasets. Using the test dataset, five, 10 or nine variables were subjected to multiple logistic regression analysis (sex was excluded [P > 0.05]). MAIN OUTCOME MEASURE: Mortality at 28 days after the first ICU day. RESULTS: Areas under the Receiver Operating Characteristic curve (AUROCs) for the test dataset in the C, P+ and P- models were 0.84, 0.89 and 0.87, respectively. Predicted mortality for the validation dataset gave Hosmer-Lemeshow chi2 values of 12.91 (P = 0.12), 10.76 (P = 0.22) and 13.52 (P = 0.1), respectively, and AUROCs of 0.84, 0.89 and 0.90, respectively. CONCLUSIONS: Our P- model is robust and does not depend on disease identification. This is an advantage, as errors can arise in coding of primary diagnoses. Our model may facilitate mortality prediction based on administrative data collected on ICU patients.