We read with interest the report by Cohen et al. (2010) dealing with the diagnosis of meningitis in Malawi, in which cryptococcal meningitis (CM) and tuberculous meningitis (TBM) were the most common specific diagnoses. The authors determined that these conditions could not be reliably distinguished on the basis of clinical and basic laboratory information alone and concluded that better diagnostic tests are needed. We agree with this conclusion, but wish to draw attention to three issues in studying diagnostic strategies in low-resource settings.
The first is the importance of selecting easily measured predictors for analysis. Of the five variables identified in the multivariate logistic regression model, only three were retained in the classification and regression tree (CART) model. The authors do not specify why the Glasgow coma score and temperature were not maintained in the algorithm, despite the fact that temperature was associated with the largest ß coefficient (ß 3.34; Odds ratio 28.01, 95% confidence interval 3.1–251.69). This may have been a result of the CART methodology itself and represents a caveat to its interpretation (Muller & Mockel 2008). Moreover, 103/263 patients with CSF abnormalities were on antiretroviral therapy (ART) at presentation, suggesting that some cases of meningitis may have been attributable to the immune reconstitution inflammatory syndrome. Unfortunately, no details are given about the duration of ART before presentation, or about whether this was analysed as a potential discriminator between CM and TBM.
Second, miscategorisation of patients affects the interpretation of diagnostic accuracy. Diagnosis of TBM is difficult in any setting, and it is likely that some of the 75 patients with ‘pyogenic’ (with a median of 81% lymphocytes in their CSF), ‘lymphocytic’, and ‘unspecified’ meningitis actually had TBM. Previous studies in sub-Saharan Africa found that the correlation between antemortem diagnosis and autopsy findings is poor (e.g. the clinical diagnosis of TB in adults had a sensitivity of 43–80% and a specificity of only 67–76%) (Rana et al. 2000; Murray et al. 2007) and that unsuspected treatable infections are not uncommon (Lucas et al. 1994). Short of the systematic implementation of autopsy studies in low-resource settings, there is a need in Africa for prevalence studies that employ all relevant gold-standard investigations – not just those that are locally available.
Third is the problem of the validation of the diagnostic strategy. Although it is not explicitly stated, it appears that the CART model was only applied to the 158 patients meeting the authors’ definition of TBM or CM, yielding a sensitivity of 92% and specificity of 58% for the diagnosis of CM. However, there is no way for a clinician to tell who these patients are at the time of clinical presentation. The model should be applied to the entire group of 563 patients presenting with suspected meningitis. Had this been performed, the CART model would have been less discriminatory still.
Hence, the need for improved diagnostics. Of all available or projected diagnostic tests to distinguish CM and TBM, assays aimed at detection of cryptococcal antigens are probably the most discriminatory single test, assuming co-infected patients are rare. Unfortunately, the current form of these assays usually requires refrigeration or freezing of reconstituted reagents, extensive sample processing, which includes boiling of samples and inactivation of competing proteins for up to one hour, and complex laboratory manipulations. There is an urgent need for simple, inexpensive and robust cryptococcal antigen detection assays, ideally using non-invasive samples such as urine.