Results
Clinical characteristics of our patient population with sepsis stratified by hospital mortality are shown in Table 1. One third of the patients died in the hospital, most presented with suspected or confirmed severe sepsis in the ED and most patients were from North America. If the patients were presented to the ED with sepsis, then more survived than died (61.3% vs 44.6%, respectively), whereas if the patients were presented to ward or ICU with sepsis, then fewer survived than died (28.9% vs 39.7% and 9.8% vs 15.7%, respectively). Approximately 90% of the patients had cardiovascular organ failure independent of mortality status, whereas with 25.1% of those who survived and 39.2% of those who died had a serum lactate more than 4 mmol/L. Pneumonia was the most common infection source followed by urinary tract infection (UTI) and abdominal infection. For those who died, renal, hepatic, hematologic, and pulmonary organ failure was higher, and they had a higher proportion of three or more baseline organ failures The proportion receiving mechanical ventilation was higher if they died.
The logistic regression model coefficients, SE, odds ratios, the 95% CI for the odds ratios, p values, and the point value for each of the predictor variables are shown in Table 2. The final multivariable model contained no significant interactions at the 0.05 level and no continuous variables. Measures of model performance for the development and validation models are given in Table 3. Both p values for the Hosmer-Lemeshow GOF statistic are greater than 0.05, suggesting that the model produces probabilities that reflect the true mortality experience of the data (i.e., well calibrated). Each model's ability to discriminate hospital mortality is considered good based on the area under the ROC curve being in the 0.70–0.80 range. The bootstrapping simulation validates that the developed model has good calibration as 4,886 of the 5,000 Hosmer-Lemeshow GOF p values (97.7%) were greater than 0.05. A sensitivity analysis where multiple imputations were used to impute the missing lactate values in the developmental dataset was run to check the missing at random assumption. The area under the ROC cures changed from 0.736 to 0.731, whereas the Hosmer-Lemeshow GOF goes from 0.373 to 0.409 verifying this assumption. Figure 1 compares the predicted and observed hospital mortality for all subjects based on the development model. These results indicate that estimated and observed hospital mortality pairs are very similar.
(Enlarge Image)
Figure 1.
Comparison of observed versus predicted hospital mortality in the deciles of predicted hospital mortality based on the developmental logistic regression model. All 23,428 observations in the dataset are presented in the figure. The solid line is the 45-degree line.
The SSS is generated by summing the individual point values (Table 2) based on all of the patients' clinical characteristics and whether they had any goal-directed therapies. The mean SSS is 56.3, whereas the median is 55 with the minimum and maximum being 0 and 130, respectively. Probability of hospital mortality based on SSS was estimated using logistic regression. The scale of SSS was determined to be linear in the logit using fractional polynomials. The coefficient for SSS is 0.0366 (95% CI, 0.0351–0.0381; p < 0.001), whereas the model intercept is –2.8334 (95% CI, –2.9279 to –2.7388). The Hosmer-Lemeshow GOF p value is 0.582 and the area under the ROC curve is 0.701, indicating good calibration and acceptable discrimination. In order to estimate the probability of hospital mortality, one would multiply the SSS by the coefficient then add the intercept. This sum would be exponentiated and then divided by the addition of one plus the expedentiated sum. Figure 2 illustrates the comparison of observed versus SSS predicted hospital mortality and closely follows the 45-degree line indicating the predicted is very similar to observed mortality over all deciles.
(Enlarge Image)
Figure 2.
Comparison of observed versus predicted hospital mortality in the deciles of predicted hospital mortality based on the Sepsis Severity Score. All 23,428 observations in the dataset are presented in the figure. The solid line is the 45-degree line.