Introduction Tuberculosis is a major challenge for antiretroviral therapy (ART) services in resource-limited settings where patients typically enrol with advanced immunodeficiency . Many patients referred for ART have a current TB diagnosis, and an additional large burden of disease is detected during pre-treatment screening –. Tuberculosis in this population is a major cause of morbidity and mortality ,– and presents a substantial hazard of nosocomial disease transmission to other patients and health care workers . These risks are heightened when patients have multidrug-resistant TB (MDR-TB) –. To address these challenges, there is a critical need in such settings for rapid, effective screening for TB and detection of drug resistance ,. Screening for TB in this patient population is difficult, however . The World Health Organization's (WHO) intensified case finding symptom screen has low specificity and misses approximately 10%–20% of cases ,. Sputum smear microscopy, the mainstay of TB diagnosis in resource-limited settings, detects as few as one in five cases when used as a screening tool pre-ART ,,. Chest radiography is costly and not widely available; interpretation is difficult, and up to one-third of culture-confirmed cases of pulmonary TB diagnosed during screening have a normal radiograph ,. Availability of culture-based diagnosis is also extremely limited in resource-limited settings because of high cost and technical complexity, and this approach often provides a diagnosis only after several weeks ,. These challenges are further compounded by the extremely limited laboratory capacity to detect drug resistance . The threat posed by MDR-TB to efforts to control TB worldwide  requires urgent improvements in diagnostic capacity. Following a large multi-country evaluation , the WHO, in December 2010, endorsed the roll-out of a novel rapid test for the investigation of patients suspected of having TB, especially in settings with a high prevalence of HIV-associated disease and/or MDR-TB . The Xpert MTB/RIF assay (Cepheid) is a fully automated molecular assay in which real-time polymerase chain reaction technology is used to simultaneously detect Mycobacterium tuberculosis and rifampicin resistance mutations in the rpoB gene ,. The cartridge-based system dispenses with the need for prior sputum processing and requires minimal laboratory expertise, and results are available in less than 2 h, permitting a specific TB diagnosis and rapid detection of rifampicin resistance. Excellent performance characteristics were observed among symptomatic adults with suspected TB in a large multi-country evaluation . These findings have been confirmed in a subsequent multi-country implementation study  and in several laboratory-based studies –. The assay has sensitivities of 98%–100% for smear-positive pulmonary TB, 57%–78% for smear-negative pulmonary TB, and 53%–81% for extrapulmonary TB when testing a variety of clinical samples ,–. Further studies are needed to examine the performance of the assay in different clinical settings, including use as a routine screening test to increase TB case detection in HIV-infected patients. We evaluated the diagnostic accuracy of the Xpert MTB/RIF assay among consecutive patients with advanced immunodeficiency being screened for TB (regardless of symptoms) prior to starting ART in a South African township with a very high burden of TB. Methods Setting The ART cohort was based in Gugulethu township, Cape Town, where the prevalence of HIV and the TB notification rate are both extremely high . Several studies reporting the burden, diagnosis, and complications of TB in this cohort have previously been published ,,,,,. National TB programme guidelines recommend investigating symptomatic adults with suspected pulmonary TB using smear microscopy of two sputum samples; in suspected “retreatment TB” cases only, culture of one sputum sample may be requested in addition . In accordance with the national ART programme guidelines, ART was provided for all patients with WHO stage 4 disease and/or blood CD4 cell counts 200 cells/µl were 28.1% (95% CI, 19.7–36.4), 19.4% (95% CI, 14.7–24.0), and 13.8% (95% CI, 10.2–17.5), respectively. In binomial regression analysis (Table 2), risk of TB was independently associated with low CD4 cell count, low body mass index, high viral load, not previously having received TB treatment, and having a positive WHO symptom screen. However, risk of TB was not associated with chronic cough of ≥2 wk duration. 10.1371/journal.pmed.1001067.t002 Table 2 Binomial regression analysis showing crude and adjusted risk ratios for the associations between risk of sputum culture-positive tuberculosis and patient characteristics. Patient Characteristics Crude Risk Ratio 95% CI p-Value Adjusted Risk Ratio 95% CI p-Value Age ≤30 y 1 Age >30 y 0.90 0.61–1.34 0.62 Male 1 Female 1.06 0.70–1.61 0.79 Body mass index 18–25 kg/m2 1 1 Body mass index 25 kg/m2 0.68 0.42–1.09 0.109 0.70 0.39–1.27 0.243 No history of previous TB treatment 1 1 History of previous TB treatment 0.68 0.41–1.13 0.14 0.50 0.26–0.96 0.036 CD4 ≥100 cells/µl 1 1 CD4 2 wk compared to 56.5% (95% CI, 41.6–71.4) among those with either no cough or cough of shorter duration (p = 0.018). Moreover, sensitivity was substantially greater in patients for whom the time to positivity of sputum samples was less than the median of 16 d (85.7%; 95% CI, 69.4–100) than in those with longer times to positivity (48.5%; 95% CI, 30.4–66.5) (p = 0.005). There was also a weak association between sensitivity and CD4 cell counts: sensitivity was 78.9% (95% CI, 58.8–99.1) in those with CD4 cell counts <100 cells/µl compared to 54.3% (95% CI, 36.9–71.6) in those with higher CD4 cell counts (p = 0.075). However, there was no association with radiographic abnormalities or with a positive WHO symptom screen. There were three patients with apparent false-positive Xpert MTB-RIF assays, giving an assay specificity of over 99.0% in each of the different analyses (Table 3). Review of the study and clinical records of these patients revealed that two of these patients had overt pulmonary and systemic symptoms suggestive of TB, and both had chest radiographs revealing parenchymal consolidation and marked hilar and paratracheal lymphadenopathy highly suggestive of TB. One of these patients was reinvestigated during routine clinical follow-up and had two positive sputum smears (2+ and 3+). Both patients received standard treatment for TB and made excellent clinical responses. The third patient had symptoms and an abnormal chest radiograph but was lost to follow-up. Use of Xpert MTB/RIF in Screening Algorithms To further explore the utility of the Xpert MTB/RIF assay, we considered clinical populations with a TB prevalence of 20%, 15%, 10%, or 5%. With an overall sensitivity of 73.3% and specificity of 99.2% (Table 3), the PPVs at these TB prevalence rates would be 95.8%, 94.2%, 91.0%, and 82.8%, respectively, and the NPVs would be 93.7%, 95.5%, 97.1%, and 98.6%, respectively. We next considered the utility of incorporating the Xpert MTB/RIF assay into different screening algorithms, examining the use of smear microscopy, symptom screening, one Xpert assay, two Xpert assays (Xpert done on a second sample if the first was negative), and sequential smear microscopy and Xpert testing (Xpert tests done if smear microscopy was negative). This was simulated for a hypothetical cohort of 1,000 patients with a TB prevalence of 20%, 15%, 10%, or 5% and assuming that 30% of cases were smear-positive. Symptom frequencies and the sensitivity and specificity of the Xpert assay as reported above were used. The yield of TB cases, the number of missed cases, and the number of Xpert tests done for each correct TB diagnosis were compared between these different screening strategies and clinical populations (Table 4). Compared to a base case scenario of smear microscopy of two sputum samples in patients with a positive WHO symptom screen, the sensitivity of algorithms incorporating the Xpert MTB/RIF assay was much greater and the corresponding number of missed diagnoses was far fewer. However, at a TB prevalence of 5%, the number of Xpert tests done per case diagnosed was high (Table 4). A strategy of sequential smear microscopy and then Xpert testing of smear-negative patients yielded the same number of diagnoses, but did not substantially reduce the number of Xpert tests per case diagnosed. 10.1371/journal.pmed.1001067.t004 Table 4 Utility of the Xpert MTB/RIF assay for tuberculosis diagnosis when incorporated into different screening algorithms and when used in hypothetical patient cohorts with a tuberculosis prevalences of 20%, 15% 10%, or 5%. Investigation Strategy Sensitivity (Percent)a Specificity (Percent) TB Prevalence 20% TB Prevalence 15% TB Prevalence 10% TB Prevalence 5% Correct TB Diagnoses Missed TB Cases Xpert Tests per TB Diagnosis Correct TB Diagnoses Missed TB Cases Xpert Tests per TB Diagnosis Correct TB Diagnoses Missed TB Cases Xpert Tests per TB Diagnosis Correct TB Diagnoses Missed TB Cases Xpert Tests per TB Diagnosis Base case screening algorithm Symptom screen + smear ×2 27.6 100.0 55.2 144.8 0 41.4 108.6 0 27.6 72.4 0 13.8 36.2 0 Using one Xpert test in algorithm Symptom screen+Xpert ×1 50.5 99.6 101 99 6.9 75.7 74.3 9.1 50.5 49.5 13.5 25.2 24.8 26.9 Symptom screen+smear ×2+Xpert ×1 50.5 99.6 101 99 6.4 75.7 74.3 8.6 50.5 49.5 13.1 25.2 24.8 26.3 Xpert ×1 for all patients 60.1 99.4 120.2 79.8 8.3 90.2 59.8 11.1 60.1 39.9 16.6 30.1 19.9 33.2 Smear ×2+Xpert ×1 for all patients 60.1 99.4 120.2 79.8 7.8 90.2 59.8 10.6 60.1 39.9 16.1 30.1 19.9 32.7 Using two Xpert tests in algorithm Symptom screen+Xpert ×2 60.6 99.4 121.2 78.8 11.1 90.9 59.1 14.7 60.6 39.4 22.1 30.2 19.8 44.4 Xpert ×2 for all patients 73.4 99.1 146.8 53.2 13.2 110.1 39.9 17.8 73.4 26.6 26.8 36.7 13.3 54.1 a Sensitivity based on the assumption that 30% of cases are sputum smear-positive. Use of symptom pre-screening limited the sensitivity of TB detection. In populations with high TB prevalence, Xpert testing of all patients regardless of symptoms increased sensitivity without substantially increasing the number of Xpert tests done per TB case diagnosed (Table 4). Compared to the strategy of doing an Xpert assay on one sputum sample from patients with a positive symptom screen, a strategy of doing two Xpert tests on all patients was associated with 22.9% higher sensitivity for TB and the fewest missed cases. Although the latter strategy would require a large absolute number of tests, at a TB prevalence of 20%, one extra TB case would be diagnosed for every additional 6.3 tests done. Detection of Rifampicin Resistance Among 81 cases of TB diagnosed, four cases had isolates resistant to rifampicin because of MDR-TB (prevalence, 4.9%; 95% CI, 1.4–12.2). Among the 445 patients (839 samples) with results of culture, drug susceptibility testing, and Xpert MTB/RIF assays all available, there were 84 isolates from 55 patients (including all four cases of MDR-TB) in which rifampicin susceptibility could be compared. Rifampicin resistance was correctly identified in all four cases of MDR-TB by the Xpert MTB/RIF assay (100% sensitivity) (Table 5). However, the Xpert MTB/RIF assay also reported rifampicin resistance in three samples from three further patients in which the isolates were reported as rifampicin susceptible using comparator assays (Table 5). A paired sputum sample was available from two of these patients and rifampicin-susceptible M. tuberculosis was reported by Xpert MTB/RIF assay in both. To resolve these discrepancies, the rpoB regions of all five isolates from these three patients were sequenced. All were found to be wild-type, confirming absence of genotypic rifampicin resistance and indicating that the three Xpert MTB/RIF assay results were false positives. All remaining patients with susceptible isolates were correctly identified as such by the assay. Thus, in a per-patient analysis, the PPV of the Xpert MTB/RIF assay for detecting rifampicin resistance was 4/7 (57%) and the specificity was 48/51 (94.1%; 95% CI, 84.8–98.8). 10.1371/journal.pmed.1001067.t005 Table 5 Comparison of results regarding drug susceptibility testing for rifampicin among paired samples from patients (n = 6) in whom rifampicin resistance was detected using one or more assays. Patient Number Sputum Smear Xpert MTB/RIF MTBDRplus on Sputum MTBDRplus on Culture Isolate MGIT Phenotypic DST rpoB Gene Sequencing Final Rifampicin Susceptibility Overall Susceptibility Pattern Concordant susceptibility results #020 NEG/NEG −/R −/− −/R −/R − Resistant MDR-TB #099 POS/POS R/R −/R R/R −/− − Resistant MDR-TB #208 NEG/NEG R/− −/− R/R R/R − Resistant MDR-TB #292 NEG/POS R/R R/− R/R R/− − Resistant MDR-TB Discordant susceptibility results #039 NEG/NEG R/S S/− S/S S/S WT/WT Susceptible Pan-susceptible #157 POS/POS R/S S/S S/S S/S WT/WT Susceptible Pan-susceptible #322 POS R − S S WT/WT Susceptible Pan-susceptible DST, drug susceptibility testing; NEG, smear-negative; POS, smear-positive; R, resistant; S, susceptible; WT, genotypically wild-type. Time to Diagnosis The median delays between sputum collection and results being available to the clinic for smear microscopy and Xpert MTB/RIF assays and positive liquid cultures were 3 d (IQR, 2–5) and 4 d (IQR, 3–6), respectively. The median delays for culture results were 12 d (IQR, 10–14) and 20 d (IQR, 17–27) for smear-positive and smear-negative disease, respectively. Cultures were incubated for 42 d before being declared negative for M. tuberculosis, with a median time to reporting of 43 d (IQR, 43–45). For the patients with confirmed MDR-TB (n = 4), the mean time to TB diagnosis and detection of rifampicin resistance was 2 d using Xpert MTB/RIF assay, 21 d using the MTBDRplus assay on a positive culture isolate, and 40 d using phenotypic drug susceptibility testing in liquid culture. Discussion A high prevalence (17.3%) of culture-proven pulmonary TB was diagnosed in this patient population, but conventional diagnostic tools widely used in resource-limited settings performed poorly. Smear microscopy detected just 28% of cases, and chest radiology was of low discriminatory value. Even using automated liquid culture as the diagnostic gold standard, diagnosis was slow, with a median delay of almost 3 wk among those with smear-negative disease. In contrast, the Xpert MTB/RIF assay was able to diagnose with extremely high specificity all cases of smear-positive TB and almost two-thirds of smear-negative cases and three-quarters of cases overall when testing two samples. Only 0.6% of test results were indeterminate. The assay also rapidly detected rifampicin resistance in all four cases of confirmed MDR-TB. However, false-positive rifampicin resistance results were also observed. The TB prevalence and associated risk factors detected in this clinical setting were similar to those previously reported from this and another ART clinic in South Africa ,,. Almost 30% of patients with CD4 cell counts <100 cells/µl had culture-proven TB, and rapid diagnosis is needed since such patients have high mortality risk ,. Only one-quarter of all TB patients reported a cough lasting ≥2 wk—a symptom screen widely used for many years to define suspected TB cases. Use of the new WHO symptom screening tool , had higher sensitivity but still would have missed 13 of the 81 TB diagnoses made in this study, suggesting the need for routine microbiological screening of all patients in this setting. We evaluated the utility of the Xpert MTB/RIF assay as a screening tool in consecutive HIV-infected adult patients enrolling for ART, excluding those who already had a TB diagnosis (approximately one-third of referrals to this cohort ). Since patients were screened regardless of the presence or absence of symptoms, our study is likely to have diagnosed TB cases at an earlier stage in the disease course than studies in which symptomatic patients were tested. In contrast, the previous Foundation for Innovative New Diagnostics multi-country evaluation  enrolled only patients with overt TB symptoms; all had a chronic cough of at least 2 wk duration and were able to produce three 1.5-ml sputum specimens. Early disease in our study would tend to be associated with lower bacillary numbers in sputum samples, as indicated by the observations that almost 70% of cases were sputum smear-negative and the prolonged median time to positivity of liquid cultures. This patient population therefore represents a major challenge for any diagnostic assay . The limits of detection of the Xpert MTB/RIF assay (95% sensitivity) defined by in vitro experiments is 131 bacilli/ml of sputum, which approaches than that of liquid culture, which falls within the range 10–100 bacilli/ml ,. In contrast, smear microscopy is able to detect only samples with more than approximately 10,000 organisms per millilitre ,. Testing a single sputum sample using Xpert MTB/RIF allowed diagnosis of all smear-positive cases regardless of smear grade; these cases pose the greatest infectious hazard within the community and health care settings. As anticipated , the sensitivity for smear-negative disease was lower than that reported in the previous multi-country evaluation  (43.3% versus 72.5% using one sputum sample; 63.3% versus 85.1% using two samples). Presence of cough of ≥2 wk was associated with much higher sensitivity for smear-negative TB, as was shorter time to culture positivity. The latter observation suggests that sensitivity was likely to have been limited by very low numbers of bacilli in sputum samples. Three patients had false-positive TB diagnoses using Xpert MTB/RIF compared to the predefined laboratory gold standard of liquid culture. However, the clinical and radiological features in these cases were highly suggestive of TB; one was confirmed as having smear-positive TB on reinvestigation, two exhibited excellent responses to TB treatment, and the third patient was lost to follow-up. These follow-up data suggest that some or all of these false-positive Xpert MTB/RIF assays may actually have been correct. The proportion of cultures lost to contamination was very low (3.1%), highlighting possible over-decontamination in the laboratory and loss of sensitivity in the culture gold standard. If this was the case, the PPV of the assay would be higher, which would increase assay utility, especially in clinical populations with lower disease prevalence. Few Xpert MTB/RIF assays were indeterminate, but the observation that three out of five of these were in culture-positive cases suggests that indeterminate results should be followed up by a repeat test. Despite only moderate sensitivity for smear-negative disease, Xpert MTB/RIF nevertheless increased overall case detection by 36% when testing one sample and by 45% when testing two samples, compared to smear microscopy. Used for baseline screening evaluation of patients enrolling in this ART service, Xpert MTB/RIF testing of a single sputum sample would detect TB in approximately 10% of the cohort, and testing two samples would detect TB in 12.5%. Thus, the assay would detect approximately one TB case for every eight patients screened, compared to one in 18 patients screened using sputum microscopy. We explored the potential impact of incorporating the assay in several screening algorithms applied to clinical populations with a range of TB prevalence rates. The likely benefits (increased TB yield) and assay costs (tests done per case diagnosed) were highly dependent on TB prevalence, and at a prevalence rate of 5%, the number of tests done per case diagnosed was high (4-fold higher than for a population with a prevalence of 20%). A strategy of screening with sputum microscopy and then testing smear-negative samples with Xpert MTB/RIF assay would result in minimal savings with regard to the number of Xpert tests done but would also result in failure to diagnose MDR-TB in highly infectious smear-positive cases. Symptom pre-screening restricted sensitivity and, at higher TB prevalence rates, did not substantially reduce the number of Xpert MTB/RIF tests done to identify one case of TB when compared to a strategy of testing all patients regardless of symptoms. Screening two samples with Xpert MTB/RIF would substantially increase the absolute number of tests done, but at high TB prevalence rates the high incremental yield may justify this approach. The number of Xpert MTB/RIF assays done might logically be stratified by CD4 cell count since this is a strong predictor of TB prevalence. For example, in high-burden settings such as South Africa, two tests might be done for those with CD4 cell count <200 cells/µl and just one test for those with higher counts. These strategies need to be evaluated by detailed cost-effectiveness analyses that take into account not simply the costs of testing but also the downstream impact on clinical outcomes and associated costs. Since the Xpert MTB/RIF instrument was based in a centralised laboratory service, with results reported via the routine laboratory system, the median time to diagnosis was similar to that of smear microscopy (4 d versus 3 d, respectively). The time to diagnosis of smear-negative disease, however, was shortened by a median of 2 wk compared to culture. Time to diagnosis and treatment would potentially be further shortened by location of the instrument in the ART clinic . The assay also has the potential to shorten the time to exclude a diagnosis of TB; this normally takes 6 wk or more via negative cultures and may lead to inappropriate delays in ART initiation. In view of the high NPV of the Xpert MTB/RIF assay in this cohort (94.8%), a negative result at baseline evaluation could provide a useful indication of a low probability of TB, increasing clinical confidence to start ART without undue delay. In cohorts with a lower prevalence of TB, the NPV would be higher, further increasing its utility in this regard. HIV-associated MDR-TB carries a high mortality risk, and nosocomial outbreaks in HIV care and treatment centres pose a grave threat to patients accessing these services ,,. Many patients with HIV-associated MDR-TB die before a diagnosis can be made ,. In this study, the Xpert MTB/RIF assay identified four patients with rifampicin-resistant isolates who had MDR-TB, greatly reducing the mean time to detection (2 d) compared to using conventional culture-based susceptibility testing (40 d) or using line probe assays on culture isolates (20 d). By accelerating diagnosis, the Xpert MTB/RIF assay has the potential to substantially reduce the risks of nosocomial transmission of MDR-TB and improve the prognosis of affected individuals. The Xpert MTB/RIF assay reported three false-positive rifampicin resistance results. The finding of discordant rifampicin susceptibility results from paired samples using the Xpert MTB/RIF assay suggests that specificity might be increased by requiring confirmation of resistance in more than one sample. While such false positives were not found in the initial multi-country evaluation , another ongoing field study sponsored by the Foundation for Innovative New Diagnostics has also detected cases, leading the manufacturer to modify the instrument software and cartridge specifications ,. With WHO approval of roll-out of this assay in December 2010, confirmation of successful reconfiguration of the test platform is urgently required. Strengths of the study include the use of a quality-assured laboratory that participated in the previous multi-country evaluation . Whereas all previously published studies have evaluated use of the assay among individuals with suspected TB ,–, this study evaluated the assay as a screening tool in unselected consecutive patients regardless of symptoms in a high-burden setting. The TB status of all patients was clearly defined based on a rigorous laboratory gold standard. Weaknesses include the fact that a small number of tests were not done because of a laboratory clerical error and that there were few cases of MDR-TB. While a similar burden of disease has been reported from an ART clinic elsewhere in South Africa , the prevalence of TB may differ in other countries, and we therefore explored utility at a range of prevalence rates. The impact of the sputum concentration procedure and of dividing the sputum pellet between three assays rather than testing unprocessed sputum was not investigated in this study, but these methods were not found to impact assay sensitivity in a previous large-scale multi-country evaluation . The usefulness of the assay as a point-of-care test was not evaluated. Further studies are needed to assess the impact of Xpert MTB/RIF screening on subsequent patient outcomes, the operational feasibility of using the assay within the clinic, and cost-effectiveness. In conclusion, when used as a routine screening test among patients with advanced immunodeficiency and high TB risk, rapid screening using the Xpert MTB/RIF assay substantially increased case detection, supporting replacement of microscopy as the initial diagnostic tool. The assay also greatly decreased the time to diagnosis of MDR-TB. Use of Xpert MTB/RIF as a screening tool might effectively reduce the risk of nosocomial MDR-TB outbreaks in HIV care and treatment settings and improve the prognosis of affected patients. However, the specificity of the assay for detecting rifampicin resistance needs to be improved to prevent overdiagnosis of rifampicin-resistant disease. Supporting Information Text S1 STARD checklist. (PDF) Click here for additional data file.