Page 129 - Hall et al (2015) Principles of Critical Care-McGraw-Hill
P. 129

CHAPTER 13: Assessment of Severity of Illness   95


                    examined 8724 critically  ill patients and reported that  crude death   are variables that influence a clinician’s decision to treat with corti-
                    rates in hospital varied more than twofold between ICUs in Britain and   costeroids, such as dose of norepinephrine being used (because clini-
                    Ireland. Application of the APACHE II equation produced an ROC   cians often use corticosteroids in patients who are “not responsive” to
                    value of 0.83 and failed to explain outcome in four ICUs. They con-  norepinephrine) and APACHE II (because perhaps sicker patients are
                    cluded that the American APACHE II equation did not fit their data   more likely to be given corticosteroids in practice). Thus, controls and
                    uniformly, and cited systematic differences in medical definitions and   cases would be matched as closely as possible according to baseline
                    diagnostic labeling, diagnostic mix, measurement of physiologic vari-  norepinephrine dose and APACHE II. The second step in matching is
                    ables, effectiveness of treatment, and differences in age-specific health   to determine from the literature and consensus opinion which variables
                    status between the two countries.                     at baseline are associated with increased risk of death (such as increased
                     The performance of APACHE III has been assessed in several   APACHE II score, age, number of organ systems failing, etc). One
                    countries  including  Brazil,   the  United  Kingdom,   Korea,   and   would want to match corticosteroid-treated cases to comparable non-
                                                                  119
                                                           118
                                        117
                    Australia.  In  most countries, the observed hospital  mortality was   corticosteroid-treated controls so that the patients are matched closely
                           120
                    significantly higher than the APACHE III predicted mortality rate. In   enough (numerically) for these variables associated with increased
                    the Australian study, when the model was corrected for hospital char-  risk of death. As a result, any differences found in hospital mortality
                    acteristics, the observed hospital mortality rate was not different. The   between corticosteroid-treated and nontreated controls can be attrib-
                    area under the ROC curve was 0.92. The APACHE III mortality model,   uted to treatment group and not to differences in baseline variables that
                    when adjusted for hospital characteristics, had good discrimination and   predict risk of death. In interim summary, for such a study of patients
                    calibration in the Australian adult ICU population.   who had septic shock, covariates that could be included in the match-
                                                                          ing algorithm would be age, APACHE II score (or SAPS II score), the
                        ■  PROPENSITY SCORING SYSTEMS AND CASE MATCHING   presence or absence of specific organ dysfunctions (eg, cardiovascular,
                      TO SIMULATE RANDOMIZED CONTROLLED TRIALS            respiratory,  renal,  and  hematologic),  norepinephrine  dose,  surgical
                                                                          status, and the site of primary infection.
                    RCTs are the “gold standard” of level I evidence for assessing efficacy of   To address the likelihood that a clinician would prescribe cortico-
                    new or controversial therapies. However, RCTs are expensive, often run   steroids, a propensity score (the likelihood of having received corti-
                    over several years, have tight inclusion and exclusion criteria, and so   costeroids given the key baseline characteristics) would be calculated
                    are sometimes less generalizable than observational studies. There are a   using covariates. The matching methodology would ensure that the
                    limited number of patients available to participate in RCTs, which lim-  three most relevant covariates, eg, age, APACHE II score (or SAPS
                    its the number of trials and hypotheses that can be tested. The science   II score), and propensity score are tightly matched between cortico-
                    and practice of critical care has advanced to address these limitations   steroid-treated  patients  and  the  matched  controls.  For  example,  one
                    of RCTs.                                              could decide that the matched patients must be within 5 years of age,
                     In recent years in critical care, investigators have used observational   within 2 points on the APACHE II score (4 points on SAPS II score),
                    cohorts and sophisticated case matching systems to simulate RCTs of   and within 0.6 standard deviations on propensity score. To control
                    interventions and drugs. The concept is to use an observational cohort   for potential changes over time in best treatment and supportive care
                    and  control  for  differences  between  control  and  treatment  groups  in   of patients with septic shock, patients could be matched according
                    the design as opposed to in the analyses. 121-127  Most often, differences   to date of enrollment in the cohort (eg, within 24 months of each other’s
                    between control and treatment groups are addressed in the adjusted   date of enrollment).
                    analyses, for example, by using logistic regression and including vari-  Another critical aspect of case-matched studies is that the matching
                    ables that differ between groups as covariates. Then the adjusted analy-  of controls and cases must be done while blinded to outcome (in this
                    ses determine whether there is still a statistically significant difference in   case hospital mortality) to minimize unintended and intended bias.
                    the outcomes of interest between treatment groups while adjusting for   A two-phase transfer of data from each center would be implemented
                    differences between groups in baseline characteristics. In case-matched   to ensure that the selection of matched control patients is imple-
                    studies, there are adjustments in design to balance control and treatment   mented in an unbiased manner. First, the database would be loaded
                    groups because a well-matched control group is obviously critical to the   with the (1) baseline variables needed for determination of eligibility
                    validity of such a nonrandomized study. 121-127       (screening for inclusion), (2)  variables  for matching, and  (3) treatment
                     Let us consider an observational cohort of patients in whom clinicians   group. Then patients would be matched (controls matched to cor-
                    have treated patients with septic shock with low-dose corticosteroids or   ticosteroid-treated patients) without knowledge of outcome. Once
                    have not treated with corticosteroids. An investigator wishes to deter-  the matching has been completed and the control patients have been
                    mine whether the cohort could be used to examine whether cortico-  identified and matched to each corticosteroid-treated case, then the
                    steroid treatment (compared to no corticosteroid treatment) decreases   database would be “locked” (ie, the controls and cases are now inextri-
                    hospital mortality.                                   cably linked together). Then, the hospital mortality outcome of each
                     The investigators would have to agree on a set of inclusion and exclu-  patient is loaded into the database. Finally, the statistical analyses are
                    sion criteria (eg, presence of two of four systemic inflammatory response   now done comparing control group to corticosteroid-treated group to
                    syndrome (SIRS) criteria, presence of infection, and presence of hypo-  determine whether there is an association of corticosteroid treatment
                    tension despite adequate fluid balance). Patients would be assessed for   with  hospital mortality.
                    eligibility according to the inclusion criteria and only those that are   There are strengths and weakness of this approach. Some of the
                    eligible would be included for the selection of matched patients. Of note,   strengths are as follows. Even though patients would not be prospec-
                    both treated and nontreated patients need to pass this inclusion and   tively recruited for such a study, the use of strict eligibility criteria would
                    exclusion screen (as in an RCT). After screening of patients according to   ensure that the validation of the treatment hypothesis (ie, treatment
                    eligibility, the full matching algorithm would be implemented.  effect of corticosteroids in septic shock) would be conducted in a well-
                     Within the screened cohort, control patients would be selected pro-  defined and relevant population of patients treated in practice. The
                    grammatically to match the corticosteroid-treated patients using an   biggest challenge in choosing appropriate patients for the control group
                    algorithm that matches on (1) baseline demographic and disease vari-  is overcoming the known patient selection bias due to lack of random-
                    ables that may have influenced clinicians’ decision to give corticoste-  ization to corticosteroid treatment. The very fact that corticosteroids
                    roids and (2) variables that are associated with risk of death (if  mortality   are not uniformly prescribed for patients who have septic shock allows
                    is the primary outcome of interest). In a design-based approach, the   this type of matched-patients study to be conducted. If corticosteroids
                    first set of variables that must be matched between controls and cases   were used most of the time in the eligible patients, then it would be very








            Section01.indd   95                                                                                        1/22/2015   9:37:29 AM
   124   125   126   127   128   129   130   131   132   133   134