We searched Medline through PubMed for articles published from January 1986 to December 2004 that reported on the diagnostic accuracy of the direct agglutination test and the rK39 immunochromatographic assay used on serum or blood samples for visceral leishmaniasis. The search terms were “visceral leishmaniasis”, “kala-azar”, “L.donovani”, “L.infantum”, and “L.chagasi” combined with “diagnostic accuracy”, “sensitivity”, “specificity”, “validation”, and “diagnostic performance”. This search generated 654 papers, which we subsequently combined with the search terms “Direct Agglutination Test or DAT” and, in a second step, with “rK39 and immunochromatographic test”, “immunochromatographic assay”, and “ICT or dipstick”. We obtained additional articles by citation tracking of review articles and original articles.
We included original studies only. Other inclusion criteria were current clinical visceral leishmaniasis as the target condition (not leishmanial infection or past visceral leishmaniasis); human participants; the absolute numbers of true positive, false negative, true negative, and false positive observations available or derivable from the data presented; and the reference classification judged adequate to correctly classify the target condition. We included studies that complied with these four criteria, independent of other characteristics of design quality and reporting. We excluded studies that evaluated the tests in patients coinfected with HIV and studies of the rapid version of the direct agglutination test (the fast agglutination screening test).6
We used a standard form to extract data on type of study (phase I, II, or III trials), Leishmania species (L donovani, L infantum, or L chagasi), country in which the study was carried out, characteristics of the participants, study design, and test results. Data were entered in a spreadsheet (Microsoft Excel).
The number and type of participants was recorded and categorised as confirmed cases or controls. Several controls were distinguished: healthy non-endemic controls (healthy people living in an area with no transmission of Leishmania and no exposure by travel); healthy endemic controls (people living in an endemic region but without signs or symptoms of leishmaniasis); controls with potentially cross reacting diseases (patients with confirmed disease that might give a false positive reaction to a serological test); and controls (patients with the same clinical syndrome as confirmed cases with visceral leishmaniasis ruled out by a confirmatory test with a high negative predictive value such as on spleen aspirate).
In many articles the numbers of true positive, false negative, true negative, and false positive observations were available. If not, we derived the numbers from the marginal totals and the reported sensitivity and specificity.
Assessment of study quality
Two independent readers assessed the papers according to the quality assessment of studies of diagnostic accuracy approach.7
Disagreements were resolved by a third reader who reassessed problematic papers. The quality assessment of studies of diagnostic accuracy tool lists 14 items that together assess the quality of a study (see tables A-C on bmj.com
). Important criteria are a representative spectrum of patients, correct classification of disease status, complete description of reference test and index test, blinding of the readers of the serological test to the result of the reference test and vice versa, and a low potential for verification bias. Such bias may exist if the decision to carry out the standard examination depends on the results of the examination under investigation—that is, if verification with parasitology occurred more often in patients with positive serology results.
We used standard formulas to calculate sensitivity, specificity, and diagnostic odds ratio for each study. Wilson's score method was used to calculate confidence intervals for sensitivities and specificities,8
and the normal approximation with continuity correction for the log of the odds ratio for confidence intervals for the diagnostic odds ratio.9
We explored the relation between sensitivity and specificity by plotting the log odds of a positive test result in diseased participants (cases) against the log odds of a negative test result in non-diseased participants (controls). Publication bias was assessed by graphing the estimated sensitivity and specificity against the number of cases and controls in each study.
Meta-analyses for sensitivity and specificity were carried out using logistic regression models accounting for overdispersion.10
This model weights each study proportional to its sample size while allowing for heterogeneity between studies by inclusion of a dispersion variable. The robustness of this statistical approach was explored by comparison with other statistical methods and simulation. The meta-analysis of the diagnostic odds ratio was carried out using the DerSimonian and Laird random effects model, with studies weighted using the Mantel-Haenszel method.11
To assess the heterogeneity of studies we carried out separate meta-analyses in subgroups stratified by study phase, sample size, study quality, geographical region, species of Leishmania
, type of direct agglutination test antigen, brand of dipstick, and type of controls.
Statistical analyses were carried out with the open source statistical language and environment R 126.96.36.199
We used the generalised linear models module (procedure “glm” using the quasibinomial distribution) to fit over-dispersed logistic regression models and the R meta-analysis library (procedure “metabin”) to fit the DerSimonian and Laird random effects model.