The search strategy and data abstraction were defined by a prospective protocol. We searched PubMed, and the Cochrane Collaboration Database through April 2006, using the following search terms: 'hemoperfusion or hemadsorption or hemodiafiltration or hemofiltration or hemodialysis' and 'polymyxin or polymyxin B or Toraymyxin or PMX-DHP or DHP-PMX' without language restrictions. We also reviewed bibliographies of retrieved articles and consulted with experts to identify relevant studies. Other methods of study identification included searching names of authors of relevant studies and contacting industry. Published English, Japanese, and Italian language full-text case series, cohort studies, and RCTs of DHP-PMX were eligible. Japanese articles were translated by a competent scientific/medical translator with a knowledge of the subject matter. To further facilitate translation, the translator was given instructions regarding the specific data being abstracted as well as specific statements of interest to the reviewers (for example, regarding randomization, blinding, and follow-up).
Prospective and retrospective observational studies, pre- and post-intervention design, and RCTs reporting original data on five or more adults treated with PMX-F for sepsis were included. Three authors reviewed all citations and abstracted data independently on a standardized form, and disagreements were resolved by discussion. Included trials had at least one of the following outcome measures: mean arterial pressure (MAP), doses of vasoactive agents, arterial partial pressure of oxygen/fraction of inspired oxygen (PaO2/FiO2) ratios, endotoxin levels, and mortality. We contacted authors and invited them to provide data for inclusion in the meta-analysis if we were unable to extract data directly from the publication or when relevant data were presented only in graphical form or only as subgroups (for example, survivors and non-survivors, by levels of Acute Physiology and Chronic Health Evaluation [APACHE] score). If the authors did not provide the data, these studies were excluded.
If multiple publications by the same investigator existed, the studies were reviewed carefully and/or the investigator was contacted to ensure that no data were analyzed in duplicate. At least three attempts were made to contact the corresponding and/or first investigator. Methods included e-mail and mailed letters. Three investigators independently assessed trial quality with the validated scale by Jadad and colleagues [26
], which measures blinding, randomization, withdrawals, and dropouts. A maximum score of 5 represents the highest quality trial.
The primary endpoints were change in MAP, use of vasoactive agents and PaO2
ratio at the end of DHP-PMX, and mortality. A secondary endpoint was the change in endotoxin levels after DHP-PMX. Assuming a standard deviation of 20 mm Hg for MAP pre- and post-PMX-F, a sample size of at least 70 patients would be needed to detect a change in MAP of at least 10 mm Hg in a paired analysis. For continuous variables such as blood pressure, data in the published studies generally were presented as a pooled summary of pre-PMX-F treatment versus post-PMX-F treatment rather than PMX-F versus conventional therapy. In many of the parallel studies, 'post-conventional therapy' values were not reported for this group. Therefore, for continuous outcomes, the effect size was the change (follow-up minus baseline) for each parameter in patients treated with PMX-F. The 'post-PMX-F' values used for the analyses were those 24 to 48 hours after the last PMX-F treatment. We combined data from parallel-designed trials with those from 'pre-post' studies in a meta-analysis using the generic inverse variance method. In both types of studies, we recorded the mean change from baseline values for the PMX group and variance estimates for this change, when reported. When these were not reported, we attempted to obtain these values or paired individual data directly from the authors. Not all investigators provided the information requested. For the studies in which these data were not available, we calculated these values as the difference between the mean 'pre-PMX-F' and 'post-PMX-F' values, and their variance estimates were derived from confidence intervals (CIs), standard deviations, and probability values reported in the manuscript [27
]. Among the studies in which 'pre-PMX-F,' 'post-PMX-F,' and change variance estimates were available, the median correlation between the two periods was 0.59 (range, 0.05 to 0.93). To be conservative, we assumed a correlation of 0.5 to impute missing change variance estimates in the primary analysis. We performed sensitivity analyses of this choice of correlation, using 0.05 as the most conservative estimate, and the results remained robust. With regard to the endpoint of mortality, because DHP-PMX is an invasive and costly procedure, we considered it acceptable as a treatment for sepsis if a 15% absolute risk reduction could be achieved. Assuming a 50% mortality in the conventional medical therapy group, an α of 0.05, and 80% power, a sample size of at least 182 subjects in each arm is needed for parallel studies. Studies were considered for inclusion in the mortality analysis if they reported mortality for a comparable patient group (for example, sepsis) in the ICU which was not treated with PMX-F. Death was determined at the end of follow-up (14 to 60 days), as available. Results for mortality were combined on the risk ratio (RR) scale. Because the random effects model incorporates statistical heterogeneity and provides a more conservative estimate of the pooled effect size than a fixed model, we present the results of all analyses according to a random model. Intertrial heterogeneity was estimated by chi-square test. Sensitivity analyses were predefined a priori
to evaluate the effects of study design, sample size, type of infection (gram-positive or -negative), imputed values for the correlation coefficient (discussed above), and center duplication. Because some investigators had more than one publication, for each endpoint we performed a sensitivity analysis in which we included only one study per investigator group, selecting the study with the largest sample size. We also assumed that the magnitude of change in certain clinical parameters would be dependent on the baseline value and performed a sensitivity analysis based on baseline blood pressure, PaO2
ratio, and endotoxin levels. Funnel plots were drawn to examine whether the smaller studies in the meta-analysis tended to show larger treatment effects, which might be due to publication bias.
Analyses were performed with Review Manager version 4.2 (RevMan; The Cochrane Collaboration 2003, Nordic Cochrane Centre, Copenhagen, Denmark). The level of statistical significance is set at a P value of less than 0.05. For continuous outcomes, the changes in the parameter (for example, MAP) are expressed in their original linear scale as a point estimate with 95% CIs and P value. For mortality, values for RR are expressed as a point estimate with 95% CIs and P value. All RRs refer to the risk for the PMX group compared with the conventional medical therapy group (labeled in graphs as 'PMX' and 'Conventional,' respectively).