Wisnivesky et al1 have recently addressed the conflicting issue of whether to place patients in respiratory isolation to prevent the nosocomial transmission of tuberculosis (TB). They provide a clinical rule derived from a case-control study that could avoid a substantial proportion of unnecessary isolation episodes.
We would like to contribute to the debate by commenting on some issues in their work. First, the authors compare patients isolated with a diagnosis of TB (cases) with patients isolated without TB (controls). Nevertheless, the true problem is not to distinguish patients with and without TB among those who have been isolated, but to distinguish them among the whole population of patients admitted to the hospital. In fact, 8 of the 73 patients with TB in this study were never isolated. It is likely that the sensitivity of the model would be lower should these 8 patients be included. Second, the study was performed in a mixed population of human immunodeficiency virus (HIV)–infected (55%) and noninfected (45%) patients, but the authors do not mention the proportion of HIV-positive and HIV-negative patients among cases and controls. It is well known that clinical and radiological presentation varies considerably between HIV-positive and HIV-negative patients,2 as well as among HIV-positive patients with different CD4 cell counts.3 Before accepting the validity of the clinical rule for both HIV-positive and HIV-negative individuals, a separate analysis of the model performance in both populations should be performed. We think that 2 separate studies in these 2 populations of patients would better address this issue. Third, it is interesting that the reviewer's final impression appears among the strongest predictors of TB. We have also found that expert clinical judgment is more reliable than prediction models in the decision of whether to place an HIV-infected patient in respiratory isolation,4 thus emphasizing the importance of clinical experience in the management of complex medical decisions. Finally, the interpretation of models derived from clinical practice, such as the models in the study by Wisnivesky et al,1 should take into account that the higher the proportion of patients who are erroneously classified, the easier it is to find models that improve clinical practice.