Statistical modeling of prognostic indices

Introduction

Severity scoring models can provide accurate outcome prediction but their performance is very influenced by variations in patient case-mix. Therefore, none of the usual scoring systems (APACHE II, SAPS II and MPM 24) fitted to this ICU: they had good discriminatory power but poor calibration. Logistic regression analysis of their variables was performed to identify the most predictive association to ICU mortality.

Methods

Data of 823 consecutive patients (pts) admitted to the ICU were prospectively collected. Pts who stayed less than 24 h at the ICU or were burn or had less than 16 years old were excluded. For pts with several admissions, only the first ICU admission was considered. The remaining 709 pts were divided in two groups: 418 (59%) pts constituted the development set and 291 (41%) pts became the validation set. After calculating the scoring indices, their variables and respective weights were separately analysed. Variables with P value <0.05 at univariate analysis were included as independent variables at logistic regression and vital status at ICU discharge was considered as dependent variable.

Results

There were 67% male and 33% female pts; median age was 46 years old, postoperative care took up 330 (46.7%) cases, of which 275 (83%) were emergency surgery. Trauma was the admission cause for 200 (28%) pts. ICU mortality rate was 25.1% and hospital mortality 33.7%. APACHE II was 16.7 ± 8.4 and SAPS II was 33.5 ± 16.5. Through statistical modeling, an hibrid model was generated, with variables and points from the three indices. With this model, the prediction obtained was: development set with discrimination ROC=0.89 and calibration goodness-of-fit C = 1.68 and validation set with ROC = 0.84 and goodness-of-fit C = 7.72.

Conclusion

Hemodynamic instability, infection, impaired renal function, respiratory failure and coma were the best predictors of death. Early identification of patients at major risk may allow treatment with more resources and interventions, in order to improve survival. Furthermore, this study shows that suitable statistical management may be useful to customize and enhance the prognostic accuracy of the currently available scoring systems.

Table 1