We evaluated incremental cost-effectiveness among diagnostic options which resulted in successful treatment of at least 2/3 (67%) of PCP patients. After excluding both strongly dominated (less effective and more expensive) and weakly dominated (equally effective but more expensive, or equal in cost but less effective) options, five procedures remained for inclusion in an incremental cost-effectiveness analysis: induced sputum with TBO, PCR, or nPCR; and expectorated sputum with nPCR or rtPCR (Table 3). Using nPCR with expectorated or induced sputum provided a relatively higher benefit at lower cost ($43 and $60 for each additional life-year gained, respectively), compared with the next-least-effective procedure (Table 3).

Variations in the cost of the diagnostic procedure had the most impact on cost per life-year gained in sensitivity analyses. Reducing the cost of the diagnostic procedures by 50% led to an approximate 50% reduction in cost per life-year gained, while doubling it led to approximately a two-fold increase in cost per life-year gained (Table S4). Modifying other factors, including the procedural sensitivity, specificity, treatment failure rates, or treatment cost had little effect on the overall cost per life-year gained. None of the analyses examined affected the relative cost-effectiveness of the diagnostic procedures with respect to each other.

Three metrics are relevant in this analysis for decision-making and policy concerning diagnostic testing for PCP: (a) proportion of PCP patients successfully treated, (b) proportion of well persons unnecessarily treated, and (c) the total diagnostic and treatment cost per life-year gained. An ideal test will maximize the first metric and minimize the second, at the smallest – and most feasible, for the implementing clinic or geographic region under consideration – value of the third. Because all laboratory-based diagnostic procedures considered in this analysis were highly specific, the effect of (b) is negligible for this analysis; thus, we presented the results as a function of (a) and (c).

Our results indicate that PCR methodologies are so sensitive that, specimen type notwithstanding, they represent the most cost-effective diagnostic options for PCP. When PCR methodologies are available, they mitigate the need for obtaining highly invasive specimens, such as BAL, which increase procedural sensitivity at substantial increases in cost. However, if both PCR and machinery for sputum induction are unavailable at a given site, the next-best option could be ES/TBO, which is relatively inexpensive and simple in terms of specimen collection and laboratory requirements for diagnosis. Although the use of chest x-ray alone for diagnosis can lead to the successful detection and treatment of high proportions of patients, the cost per life-year gained exceeds that of other equally-sensitive or more sensitive methods for diagnosing disease.

In general, the decision about which test is most useful in a given region will depend on the estimated prevalence of PCP among persons tested, local technical capacity, and available financial resources. Individual patient characteristics may affect decision-making, too; in an already-intubated patient, a BAL will be meaningfully cheaper than it would be among non-intubated patients, making the increased sensitivity in specimen collection more economical as well as practical (as an intubated patient or an infant will be unable to produce sputum). In addition, BAL might facilitate the detection of other respiratory pathogens besides Pneumocystis, such as TB or staphylococcus. Similarly, chest x-rays can provide information beyond the ability to evaluate a patient for signs consistent with PCP. However, for the diagnosis of PCP, providing a patient can produce sputum, the model presented suggests that there is little added value in carrying out a BAL over an induced sputum procedure.

There are several limitations to this analysis. First, data were not available on the sensitivity and specificity of all diagnostic procedures, creating a need to estimate some values. Even for procedures for which data were available, the degree of experience of an administering clinician or laboratory technician could affect the test's sensitivity or specificity. Second, indirect costs are not included in the model. The buildings, equipment, and technical know-how needed to carry out more advanced molecular diagnostics such as PCR are not currently in place in all countries. Where this capacity does exist, it may be unevenly distributed geographically and might not be accompanied by appropriate quality assurance measures. Start-up costs to implement these technologies could be prohibitive for some low-income countries, and in these areas comparisons of the cost-effectiveness of the various staining methods might be more useful than considerations about which PCR methodology is optimal. Finally, we did not account for differing diagnostic or treatment costs in different countries or among different patient groups, which could affect overall cost or cost-effectiveness of different diagnostic options. However, it is worth noting that, although the costs of all procedures might differ by country, the relative cost of procedures is unlikely to differ greatly.

Existing international guidelines call for CTX prophylaxis of PCP in patients whose CD4+ T cell counts drop below 350 cells/mm [3]. CTX is also considered the treatment of choice for PCP. Because diagnostics are not available throughout much of the developing world and because CTX is relatively inexpensive and can be effective in the treatment of other respiratory pathogens in addition to PCP [88], [89], clinicians in developing countries frequently diagnose symptomatic patients empirically for PCP. While this method may capture a high proportion of patients with disease, it is highly nonspecific and thus may result in many patients without PCP being treated for the disease. Why should we consider this important? First, CTX has known toxic side effects: HIV-infected patients in particular are at risk of adverse reactions to CTX, including cutaneous reactions [90], fever, neutropenia, thrombocytopenia, transaminase elevation [91], meningitis [92] and anaphylaxis [93]. Second, non-judicious use of antimicrobials has long been recognized as a precursor to increased drug-resistance for a broad spectrum of pathogens. While reports of Pneumocystis resistance to CTX are infrequent, some do exist, and concern about resistance is increasing [94], [95], [96]. Perhaps as importantly, treatment for other pathogens, including S. pneumoniae, malaria, Salmonella spp, Staphylococcus aureus, and Escherichia coli, involve the use of CTX, and for these pathogens reports of CTX drug-resistance are common [97], [98], [99], [100], [101], [102], [103].

The cost-effectiveness of diagnostic testing improves in areas of higher disease prevalence; testing might become prohibitively expensive in areas with very low prevalence of disease. However, a diagnostic protocol that might seem financially unfeasible for certain regions might be more feasible than suspected if the prevalence of disease can be increased among the patients selected for testing. One way to optimize test utility is to use a clinical algorithm that improves the pre-test probability without incurring substantial numbers of false negatives. Although no such algorithm has been formally defined for PCP, clinical differences do exist between HIV-infected patients with PCP compared with other pneumonias; PCP patients have a more subacute onset of disease, ground glass infiltrates on CXR [104], lower oxygen saturation, lower CD4 cell counts, greater weight loss, more cyanosis, more severe dyspnea, and higher respiratory rates than non-PCP patients [15], [105]. Utilizing one or a combination of these metrics might be useful for increasing the prevalence of PCP among the population to be tested (e.g., increasing the pre-test probability), provided it did not miss substantial numbers of patients with PCP.

For any disease, when the cost of diagnosis exceeds the cost of treatment (such as with PCP), the cost-effectiveness of empiric diagnosis and treatment is directly proportional to the gap between the diagnostic and treatment costs; thus, when treatment costs are very low, it's nearly always more cost-effective to diagnose and treat patients empirically. In addition, because international guidelines call for at-risk patients to be on ART and CTX prophylaxis, the occurrence of PCP in a patient likely represents a failure of the local health system to provide sufficient opportunities for HIV patient care and treatment, an inability by the treating clinic to meet these standards, an inability by the patient to adhere to the recommended treatment regimen, or drug failure. Thus, one could argue that efforts should be focused on improving access to care for HIV patients or adherence to the standards laid out in international guidelines with respect to ART and CTX treatment, rather than on diagnosing the precise etiology of infections that could otherwise have been prevented. This is a valid argument and such efforts should be supported. However, given the suboptimal conditions that currently exist with respect to meeting these guidelines, there are benefits to accurate diagnosis, including improvements in the understanding of the true prevalence of disease, which is worthwhile for the purposes of prevention, control, and allocation of resources. This analysis is not intended to discourage PCP prophylaxis or diagnosis and treatment among symptomatic patients in the absence of a laboratory-based diagnosis, but rather to provide a basis for decisions on diagnostics for PCP, should an institution desire to implement diagnostic procedures. For these institutions, particularly in situations of high disease prevalence, we demonstrate that the elevated sensitivity and specificity of diagnosis enabled by the use of PCR technologies could justify the additional costs of obtaining and using them. A rough calculation demonstrates the power of replacing microscope-based technologies with PCR technologies for the diagnosis of PCP: in South Africa, the adult HIV infection rate is reported at 20%, with an estimated half a million new infections [106] and approximately 250,000 persons starting ART each year [107]. Assuming a PCP prevalence of ∼20% among HIV-infected persons starting ART (PCP prevalence among HIV-infected children is reported to be as high as 52% in South Africa [7], [8], [18]), approximately 50,000 persons would start ART with PCP. Using PCR technologies with expectorated sputum could result in the successful treatment of between 8,350 and 27,650 more PCP infections (i.e., result in 8,350–27,650 more life-years gained) than diagnosis with non-molecular-based technologies (or empiric diagnosis and treatment).

Diagnostic procedural decisions cannot, in practice, be simplified to numbers alone. Assuming clinicians were aware of the diagnostic qualities of each test, they could make decisions outside of the framework presented here, such as conducting sequential tests (for example, a highly sensitive test followed by a highly specific test) for diagnostic purposes. In addition, we realize that most clinicians do not have an array of diagnostic options at hand, and if a diagnostic protocol is to be implemented, it will be done at a clinic, hospital, or regional level. However, examples of molecular diagnostic technologies in resource-limited settings are increasingly reported, for example with tuberculosis diagnosis [108], [109], [110], [111]. The fixed costs associated with building structures and capacity to carry out these technologies will decrease in proportion to their utility as these technologies become cheaper and address to increasing arrays of pathogens. As more patients develop PCP and concerns about CTX drug resistance grow, it is worth considering whether changes in existing diagnostic paradigms are warranted for PCP. Choosing suboptimal diagnostic methodologies – or no laboratory-based methodologies at all, for empirically-diagnosed disease - for a treatable infection may no longer be justifiable, particularly in high-prevalence areas. In recognition of the enormous increases in diagnostic sensitivity available with more technologically complex procedures such as PCR, we would encourage policymakers, particularly those in regions where disease prevalence is high among the population tested, to consider prioritizing the development of the skills and infrastructure necessary to support improved diagnostic methods. It is our hope that this analysis can serve as a guide to help clinicians or policymakers make decisions about the best use of limited resources.