Increased mortality attributed to Chagas disease: a systematic review and meta-analysis

The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines
and checklist was used to ensure inclusion of relevant information in the analysis
21] (Additional file 1).

Search strategies

Searches were conducted in PubMed (online version of Index Medicus, produced by the
USA National Library of Medicine, NLM); MEDLINE (a subset of PubMed (~98 %) made available
by NLM); EMBASE (Excerpta Medica dataBASE), Web of Science (Core Collection) and LILACS
(Latin American and Caribbean Health database), without time filters until the 31
^st
of September 2015. The search algorithm combined four search terms to represent the
grouping of the concepts most relevant to the question under scrutiny: 1) Chagas disease,
2) mortality, 3) progression, and 4) survival analysis. This search algorithm was
applied to each database to maintain consistency in the results generated. The full
search terms for individual databases are available in Additional file 2: Table S1. All the titles and abstracts were assessed by two independent investigators
(ZMC and OO), eliminating studies that did not meet the inclusion criteria: i. cohort
studies, ii) comparing Chagas and non-Chagas patients, iii) with follow-up for more
than one year. Disagreements were resolved by consensus, and in the case of persistent
disagreement, the full text of the article was examined. References cited in the selected
papers were inspected and if appropriate included as secondary searches.

Data extraction

Each paper that was selected for analysis of the full text was reviewed carefully
and the relevant information was extracted. In some instances information was extracted
from available data tables or figures, where values were not explicitly mentioned
in the text. A data extraction table was designed to obtain information from each
eligible study. The following items were included: first author; year of publication;
year of study; location of the study; study design; sample size; proportion of men
in the study population; age group; mean/median age of study participants; number
of deaths; years of follow up; number of persons-year of follow up; loss to follow
up (drop-out rate); clinical classification (severe, moderate, asymptomatic); reported
effect size and corresponding adjustments.

In order to obtain results accounting for severity of symptoms, the data were extracted
and classified according to the clinical severity reported in each study, as follows:

Severe stage: this stage included patients with cardiac complications, attending
health facilities and usually classified according to the New York Heart Association
Functional Classification (NYHA) III and IV. Also one study 22] which included only a population under resynchronization therapy was considered in
this stage.

Moderate stage: this included populations mostly classified according to NYHA I and
II criteria.

Asymptomatic/general population: this category included both asymptomatic populations?mainly
from population studies?and also individuals with minimal electrocardiogram (ECG)
damage or without report of deleterious ejection fraction.

All stages: this category included studies in which several clinical stages were
used in comparison to clinically similar but uninfected controls.

Quality assessment

The Newcastle-Ottawa Scale (NOS) was used to assess the risk of bias of the studies
included in this review in a standardised manner, as this metric is easy to interpret
and is recommended for quality assessment by the Cochrane Collaboration 23]. The NOS scale assesses each study on three components, namely, the selection of
the study population, a valuation of comparability of the study groups, and an assessment
of the outcome of interest. Each study is scored for each component by the award of
“stars”. The checklist, amendments made to the original scale and details on the assessment
for each study are presented in Additional file 2: Table S2. The critical appraisal of the studies was conducted following the data
extraction process. Three levels of quality were considered: low, moderate and high.
Due to the small number of studies identified, studies were not excluded based on
quality assessment. Nevertheless, a separate analysis was done only using papers deemed
as of “high quality”.

Statistical analysis

Studies were required to report hazard ratios (HRs), relative risk ratios (RRs), odds
ratios (ORs) and their 95 % confidence intervals (CIs) or to provide adequate data
to allow the 95 % CI to be calculated. Because not all studies reported deaths in
a uniform manner, the analysis is based on all-cause mortality, cardiac death, heart
transplant or death due to stroke. For quantitative analysis, studies were included
if enough information was provided to estimate crude RRs.

Selected studies differed substantially in terms of sample size, study location and
clinical characteristics; therefore, heterogeneity in mortality rates was potentially
important. Thus, a random-effects model was used to test differences in rates of mortality
between chagasic and non-chagasic populations. For the random-effects model, tau-squared
(?²
) was presented as a measure of the between-study variance. For comparison, results
using a fixed-effects model are also presented (Additional file 2: Figure S1).

Heterogeneity among studies was measured using Cochran’s Q test and I
²
statistic. Cochran’s Q is computed by summing the squared deviations of each study’s
estimate from the overall estimate, weighting each study’s contribution. The p-values
for this test are obtained by comparing the Q statistic to a chi square distribution
with k–1 degrees of freedom (df) (where k is the number of studies). The I
²
statistic measures the degree of inconsistency in the studies’ results. Formally,
I
²
?=?100 %?×?(Q–df)/Q, measuring the percentage of variation across studies that is
due to heterogeneity rather than to chance 24].

To explore further the source of potential heterogeneity in mortality between studies,
we used meta-regression techniques to formally identify potential covariates of the
estimated effect on mortality rates 25], 26]. Covariates tested included clinical characteristics (as defined above), starting
year of the study, sex (as proportion of males), and location of study (country).

We explored publication bias by drawing funnel plots, enabling quantification of bias
using Egger’s regression asymmetry test 27]. Interpretation of funnel plots is facilitated by inclusion of diagonal lines representing
the 95 % confidence limits around the summary effect. In the absence of heterogeneity,
95 % of the studies should lie within the funnel defined by these lines (because these
are not strictly speaking 95 % limits, they are referred to as “pseudo 95 % confidence
limits”) 28]. A trim-and-fill technique (aimed both at identifying and correcting funnel plot
asymmetry) was then used to re-estimate excess mortality correcting for publication
bias (i.e., by incorporating the hypothetically missing studies) 29].

Finally, sensitivity analyses were also performed by 1) sequentially removing one
study at a time and re-evaluating the model to explore the impact of potential outliers
on estimates of excess mortality, and 2) restricting the analysis to ‘high quality’
papers.

The crude mortality rates were calculated for each clinical group and RR values were
used for meta-analysis. Annual mortality rates (AMR) are reported (unless otherwise
stated) per person per year. The Attributable Risk Percent (ARP) was used to estimate
excess mortality above background mortality rate, as (RR???1)/RR expressed in percent.

All analyses were performed using Stata 13.1 (StataCorp, College Station, TX).