Procalcitonin as a diagnostic marker for sepsis

Abstract

We appreciate the comments of Michael Osthoff and Damon Eisen about a lack of diagnostic accuracy of procalcitonin in our metaanalysis, the need for serial testing, and their suggested treatment decisions. However, previous metaanalyses on the diagnostic accuracy of procalcitonin measurements in patients with severe sepsis in intensive-care units were limited by selected populations, and were biased by the choice of gold standard for the definition of sepsis. We noted a mean sensitivity of 0·77 (95% CI 0·72–0·81) and specifi city of 0·79 (0·74–0·84) and an area under the receiver operating characteristic curve of 0·85 (0·81–0·88). By contrast, Tang and colleagues reported a mean sensitivity and specificity of 0·71 (0·67–0·76), and an area under the curve of 0·78 (0·73–0·83). We agreed that these results are far from ideal, and concluded that results must be interpreted carefully for clinical and microbiological assessment and have to be re-evaluated accordingly. However, in the daily practice of an intensive-care unit, accuracy of clinical confi rmation of the diagnosis alone is probably much worse. We also agree that the time course is more relevant than are the initial procalcitonin concentrations in patients in intensive-care units, and that higher cutoff ranges than are used in other settings should be considered. However, we strongly argue against any biomarker-guided approach to initiate treatment in sepsis, in situations in which, among others, broad-spectrum antibiotic treatment has to be given immediately. Treatment decisions and efficacy analyses were not one of the objectives of our Article. However, we investigated the effect of procalcitonin-guided treatment compared with standard care in another meta-analysis of seven studies with 1075 patients with severe sepsis. Findings showed that procalcitonin concentrations might be helpful to guide antimicrobial treatment through a reduction in duration of antimicrobial treatment without an obvious increase in mortality. Importantly, treatment algorithms and cutoffs differ substantially and should be clarified in future studies before recommendations are made. We also thank Gerta Rucker and Martin Schumacher for their interest in our work and for their thoughtful reanalysis of our data. We agree that the handling of diff erent cutoff values is a difficult problem in diagnostic meta-analyses. This issue is due to the fact that often only one pair of sensitivities and specificities are given for a particular study. Thus, the approach of Rucker and Schumacher helps to deal with this problem and is facilitated by the additional assumption that the authors of the individual studies report a result on the basis of an optimum cutoff value. More specifi cally, the overall summary receiver operating characteristic curve in this setting is obtained as a weighted sum of study-specific receiver operating characteristic curves in the logit space. However, we are not entirely convinced that these results are better than are those obtained with the bivariate generalised linear mixed model that we used. This assertion is based on two arguments. First, the approach of Rucker and Schumacher is a fi xed-eff ects model that ignores heterogeneity between studies. Our analysis revealed substantial heterogeneity between studies, which should be taken into account. Second, the proposed method works in logit space. Simulation studies have shown that an exact binomial generalised linear mixed model has better performance compared with logit transformed data. Thus, a comparison of methods by simulation studies would be interesting and seems necessary. We are not entirely convinced that the observed lower values of sensitivity and specifi city lead to a strong argument against the use of procalcitonin as a diagnostic marker for sepsis. Furthermore, we would like to prompt authors of diagnostic studies to report sensitivities and specifi cities for more than one cutoff value, since this knowledge would allow a better consideration of cutoff values.