Avaliação do desempenho, revisão e extensão do escore prognóstico de infecção do sítio cirúrgico do sistema NNIS (National Nosocomial InfectionsSurveillance) em hospitais brasileiros

Abstract

Introduction: the NNIS (National Nosocomial Infections Surveillance) surgical-site infection (SSI) risk index is the SSI risk-adjustment score most widely used worldwide, but its performance is highly variable across operative procedures categories. Objectives: to assessthe benefit of using, for each NNIS risk index variable, procedure-specific alternative cut-off points, to better reflect specific attributes of the samples. Also, to evaluate the advantages of extending the SSI risk prediction models by incorporating a postdischarge surveillanceindicator. Method: an open-label, retrospective cohort study was conducted. Consecutive inpatients operated on between January 1993 and May 2006 at five private, nonuniversity healthcare facilities in Belo Horizonte and Contagem were included. In-hospital SSIsurveillance was accomplished by means of SSI clues and direct inspection of the wounds. Out-of-hospital surveillance was conducted by patient telephone surveys, 30 days after the surgery. All analyses were done within specific NNIS operative procedure categories: othergenitourinary (n = 20723), other integumentary system (n = 12408), other musculoskeletal (n = 15714) and abdominal hysterectomy (n = 11847). Development and validation samples were defined nonrandomly. In the development samples, alternative cut-off points for each NNIS risk index variable were defined using density histograms, contingency tablesand decision tree analysis. Alternative SSI prognostic scores were then constructed using logistic regression for the selection and weighting of covariates: i) the alternative NNIS scores were made up of NNIS risk index covariates and cut-off points, but used locallyderivedSSI risk strata and SSI rates; ii) the alternative 1 scores were made up of NNIS risk index covariates but used procedure-specific alternative cut-off points; iii) the alternative 2 scores extended the alternative 1 scores, by incorporating a postdischarge surveillance indicator. Sensitivity analysis of logistic regression coefficients was conducted by comparingthree estimation methods (asymptotic, exact and bias-corrected asymptotic). The NNIS risk index and the alternative indexes were then applied to the validation samples and their performance was compared using measures of calibration (Cox calibration regression),discrimination (area under the receiver operating characteristic curve) and overall performance (ordinal association, trend across ordered groups, Brier score, and model 2). Results: the NNIS risk index showed poor overall performance, with low discrimination, inadequate calibration and predictions with high variability. The alternative NNIS scoresperformed variably when compared to the NNIS risk index, depending on the performance measure and the operative procedure considered. The most consistent advantage was regarding the overall calibration and the prevalence and dispersion components of calibration. Alternative 1 scores performed slightly better than the NNIS risk index for mostprocedures and measures analyzed, mainly in terms of calibration. With few exceptions, alternative 2 scores performed clearly better than the NNIS risk index, irrespective of the measure or operative procedure considered. Conclusions: these data suggest that the use oflocally-derived SSI risk strata and SSI rates improves the NNIS risk index calibration. The use of alternative cut-off points for the NNIS risk index covariates can further improve the specification of the intrinsic SSI risk component. Finally, controlling for incomplete postdischarge SSI surveillance provided consistently more accurate SSI risk-adjustment.