Objective: As the Surviving Sepsis Campaign was assessing patient-level data over multiple countries, we sought to evaluate the use of a pragmatic and parsimonious severity-of-illness scoring system for patients with sepsis in an attempt to provide appropriate comparisons with practical application.
Design: Prospective, observational evaluation.
Patients: Data from 23,438 patients with suspected or confirmed sepsis from 218 hospitals in 18 countries were evaluated.
Setting: This analysis was conducted on prospective data submitted to a database from January 2005 through March 2010.
Interventions: None.
Measurements and main results: Maximum likelihood logistic regression was used to estimate model coefficients, and these were then used to develop a Sepsis Severity Score. The probability of hospital mortality was estimated using the Sepsis Severity Score as the sole variable in a logistic regression model. Univariable logistic regression determined which variables were included in the multivariable predictor model. The scale of continuous variables was assessed using fractional polynomials. Two-way interactions between variables were considered for model inclusion if the interaction p value is less than 0.05. The prediction model was developed based on randomly selecting 90% of available patients and was validated on the remaining 10%, as well as by using a bootstrapping technique. The p values for the Hosmer-Lemeshow goodnessof-fit statistic in the developmental and validation datasets were considerably greater than 0.05, suggesting good calibration. Development and validation areas under the receiver operator curve curves were 0.736 and 0.748, respectively. Observed and estimated probabilities of hospital mortality for the total population were both 0.334. The validation and the developmental datasets were gradually compared over deciles of predicted mortality and found to be very similar.
Conclusion: The Sepsis Severity Score accurately estimated the probability of hospital mortality in severe sepsis and septic shock patients. It performed well with respect to calibration and discrimination, which remained consistent over deciles. It functioned well over international geographic regions. This robust, population-specific evaluation of international severe sepsis patients provides an effective and accurate mortality estimate allowing for appropriate quality comparisons with practical clinical and research application.