The impact of the roll-out of rapid molecular diagnostic testing for tuberculosis on empirical treatment in Cape Town, South Africa

Abstract Objective To investigate the impact of introducing a rapid test as the first-line diagnostic test for drug-sensitive tuberculosis in Cape Town, South Africa. Methods Xpert® MTB/RIF (Xpert®), an automated polymerase-chain-reaction-based assay, was rolled out between 2011 and 2013. Data were available on 102 007 adults treated for pulmonary tuberculosis between 2010 and 2014. Tuberculosis notification rates per 100 000 population were calculated for each calendar year and for each year relative to the test roll-out locally, overall and by bacteriological confirmation. Empirical treatment was defined as treatment given without bacteriological confirmation by Xpert®, sputum smear microscopy or sputum culture. Findings Between 2010 and 2014, the proportion of human immunodeficiency virus (HIV)-negative patients treated empirically for tuberculosis declined from 23% (2445/10 643) to 11% (1149/10 089); in HIV-positive patients, it declined from 42% (4229/9985) to 27% (2364/8823). The overall tuberculosis notification rate decreased by 12% and 19% among HIV-negative and HIV-positive patients, respectively; the rate of bacteriologically confirmed cases increased by 1% and 3%, respectively; and the rate of empirical treatment decreased by 56% and 49%, respectively. These changes occurred gradually following the test’s introduction and stabilized after 3 years. Conclusion Roll-out of the rapid test in a setting with a high prevalence of pulmonary tuberculosis and HIV infection was associated with a halving of empirical treatment that occurred gradually after the test’s introduction, possibly reflecting the time needed for full implementation. More than a quarter of HIV-positive patients with tuberculosis were still treated empirically, highlighting the diagnostic challenge in these patients.


Introduction
Sputum smear microscopy is traditionally the first-line diagnostic test for tuberculosis in countries without routine access to the gold standard: sputum culture. This approach is limited by low sensitivity, particularly among patients who test positive for the human immunodeficiency virus (HIV), and is associated with diagnostic delays, underdiagnosis and empirical treatment. 1 The Xpert® MTB/RIF (Xpert®) test (Cepheid, Sunnyvale, United States of America) is an automated, cartridge-based, rapid molecular diagnostic test for Mycobacterium tuberculosis and its resistance to rifampicin. 2 The test detects the rpoB gene of M. tuberculosis, including mutations that encode rifampicin resistance, using a realtime polymerase chain reaction and takes less than 2 hours. Because the test has higher sensitivity than smear microscopy (88% versus 65%, respectively), 3 is rapid and has the ability to detect rifampicin resistance immediately, the World Health Organization (WHO) endorsed its use in resource-constrained settings in December 2010. 4 By the end of 2014, concessionary pricing had led to widespread roll-out of the test in lowerincome countries. 5 In 2013, WHO identified understanding the impact of the Xpert® test on individual and public health outcomes as one of the top 10 research areas in tuberculosis. 6 Modelling studies indicated the test would increase tuberculosis case-finding and that the resulting earlier treatment would improve outcomes, leading eventually to reductions in tuberculosis incidence and mortality. [7][8][9] However, the four large randomized trials published to date failed to document these reductions. [10][11][12][13] This failure may have been due to empirical treatment being replaced by bacteriologically confirmed treatment rather than to more patients being identified. 14 Subsequently, when one of the original modelling papers was modified to align its results with one of the trials, the predicted decline in tuberculosis incidence decreased from 6% to 1.6%. 10,15 Modelling the effect of the test in India produced similar results. 16 There is a need for more data on the impact of the Xpert® test in practice. In South Africa, roll-out of this test as the primary test in a new tuberculosis diagnostic algorithm started in March 2011 -it was completed in Cape Town in February 2013. 17 In this study, we evaluated the impact of the test roll-out in Cape Town on the diagnosis of patients with drug-sensitive tuberculosis, stratified by HIV status. We also analysed associated changes in the proportion of notified tuberculosis cases that were confirmed bacteriologically and examined risk factors for empirical treatment. Finally, we determined whether the changes observed increased with time following the introduction of the test.

Methods
The estimated population of Cape Town in 2011 was 3.7 million. 18 The diagnosis and treatment of tuberculosis in the city was provided by 101 government primary-care clinics; tuberculosis treatment in private clinics was infrequent. 19 The Objective To investigate the impact of introducing a rapid test as the first-line diagnostic test for drug-sensitive tuberculosis in Cape Town, South Africa. Methods Xpert® MTB/RIF (Xpert®), an automated polymerase-chain-reaction-based assay, was rolled out between 2011 and 2013. Data were available on 102 007 adults treated for pulmonary tuberculosis between 2010 and 2014. Tuberculosis notification rates per 100 000 population were calculated for each calendar year and for each year relative to the test roll-out locally, overall and by bacteriological confirmation. Empirical treatment was defined as treatment given without bacteriological confirmation by Xpert®, sputum smear microscopy or sputum culture. Findings Between 2010 and 2014, the proportion of human immunodeficiency virus (HIV)-negative patients treated empirically for tuberculosis declined from 23% (2445/10 643) to 11% (1149/10 089); in HIV-positive patients, it declined from 42% (4229/9985) to 27% (2364/8823). The overall tuberculosis notification rate decreased by 12% and 19% among HIV-negative and HIV-positive patients, respectively; the rate of bacteriologically confirmed cases increased by 1% and 3%, respectively; and the rate of empirical treatment decreased by 56% and 49%, respectively. These changes occurred gradually following the test's introduction and stabilized after 3 years. Conclusion Roll-out of the rapid test in a setting with a high prevalence of pulmonary tuberculosis and HIV infection was associated with a halving of empirical treatment that occurred gradually after the test's introduction, possibly reflecting the time needed for full implementation. More than a quarter of HIV-positive patients with tuberculosis were still treated empirically, highlighting the diagnostic challenge in these patients.
diagnosis of pulmonary tuberculosis was generally based on sputum smear microscopy, with sputum culture reserved for patients who remained symptomatic despite negative microscopy findings or who were being retreated. At all clinics, chest X-ray facilities were available, either on-site or through referral. Although the empirical treatment of tuberculosis based on symptoms and chest X-ray findings alone was discouraged, it was an accepted practice for patients who remained symptomatic despite negative microbiological findings.
The Xpert® test machines were installed in all laboratories in Cape Town between August 2011 and February 2013 and use of the test as the primary diagnostic test in the tuberculosis diagnostic algorithm for Western Cape Province was endorsed in a circular to all primary health-care clinics in January 2013. 17 Two sputum samples were collected and submitted simultaneously to a laboratory -one was for the rapid test. If a rifampicin-sensitive M. tuberculosis strain was detected, the second sample was used to determine the sputum smear status pretreatment and, thereby, helps identify smear conversion during follow-up. If a rifampicin-resistant strain was detected, the second sample was used for culture and for testing drug sensitivity. If the first sample from an HIV-infected patient tested negative, the second was used for culture. If the first sample from an HIV-negative patient tested negative, the second was discarded because of the test's higher sensitivity in these patients. When the test was introduced, the treatment regimen for previously treated patients was changed to that for new patients because rifampicin resistance could then be identified before treatment. 17 Our population-based study covered 2010 to 2014: 2010 was the last full calendar year before the test roll-out began in Cape Town and 2014 was the first full calendar year after roll-out had been completed. We included all pulmonary tuberculosis patients aged 15 years or older who started treatment during that period. To avoid duplication, we excluded patients transferred between subdistricts. We used anonymized data from the City of Cape Town electronic tuberculosis register on the patients' characteristics, microbiological test results, chest X-ray results, treatment initiation dates and treatment outcomes. Patients with drug-resistant tuberculosis were entered into a separate register and were not included in our evaluation.
We defined the primary method of diagnosis as either: (i) the rapid test; (ii) sputum smear microscopy; or (iii) sputum culture -if more than one test was positive, the primary method was the first test in this sequence. We defined bacteriological confirmation of infection as a positive result to one of these tests. Empirical treatment was defined as treatment given when no test was positive or no test was performed. In addition, patients who tested positive on sputum culture and negative on, or did not undergo, other tests were regarded as having started empirical treatment if their sputum sample was sent for culture after treatment initiation or up to 6 days before initiation (assuming that 7 days was the minimum time required for a positive sputum culture result). 20 Treatment followed national guidelines and did not differ by method of diagnosis. 21 Two definitions of time were used: the calendar year and the year relative to the time when the test was introduced in the diagnosing clinic. For the latter year, we identified the date on which the first test result was recorded for each clinic and used this date as the roll-out date for that clinic. Patients were then assigned to a year depending on when their treatment started relative to the test roll-out in their clinic: patients who initiated treatment in the 365 days before the roll-out date were assigned to year 0 and those who started treatment 1096 to 1460 days after the roll-out date were assigned to year 4; other patients were assigned to intermediate years accordingly (Fig. 1). The patient was defined as HIV-positive at the time of tuberculosis treatment if the tuberculosis register recorded: (i) a positive HIV serological test result; (ii) treatment with antiretroviral therapy or co-trimoxazole; or (iii) a CD4+ T-lymphocyte (CD4+ cell) count. The patient was HIV-negative if a negative HIV serological test result had been recorded. All other patients were regarded as having an unknown HIV status.

Statistical analysis
We used descriptive statistics to present data on patients' demographic and clinical characteristics in each calendar year and on the primary method of diagnosis in each calendar year and in each year relative to the test roll-out. We calculated annual population disease rates by dividing the total number of bacteriologically confirmed and empirically treated pulmonary tuberculosis patients aged 15 years or older in a year by the mid-year estimate of the adult population in the study area. 22 Rates were also stratified by bacteriological confirmation and HIV status. The size of the HIV-negative and HIV-positive adult population was derived using annual HIV prevalence estimates from the Actuarial Society of South Africa Western Cape AIDS demographic model. 23 We also calculated population disease rates relative to the year of test roll-out, overall and stratified by bacteriological  Rapid diagnostic test roll-out, South Africa Sabine Hermans et al.
confirmation. We estimated the population size for each year relative to test roll-out as follows, taking year 2 as an example: we calculated the proportion of all patients in each calendar year who were in year 2 of roll-out ( Fig. 1) and multiplied this proportion by the estimated population for the corresponding calendar year. We then summed these estimates for all calendar years, which gave us the total estimated population for year 2 of roll-out. Factors associated with empirical tuberculosis treatment, both overall and stratified by HIV status, were identified by multivariable logistic regression analysis. A priori risk factors included age, sex, HIV status, CD4+ cell count at the start of tuberculosis treatment, history of tuberculosis treatment, calendar year and year relative to test roll-out. Age, calendar year and year relative to test roll-out were included as either continuous or categorical variables based on the results of tests for departure from linearity. Because of the collinearity between our two-time variables, we used two separate multivariable logistic regression models -one included the calendar year and the other included the year relative to test roll-out. We accounted for clustering at the clinic level by calculating robust standard errors. In addition, a sensitivity analysis was performed using random effects models that adjusted for clustering at the clinic level. We tested for changes over time in the odds of empirical treatment in years 2, 3 and 4 of test roll-out using a model that included only those years. All analyses were performed using Stata/IC version 13.0 (StataCorp. LP, College Station, USA) and Excel 2013 (Microsoft Corporation, Redmond, USA). The study was approved by the Human Research Ethics Committee at the University of Cape Town and by the City of Cape Town Health Department.

Results
In 2010, 21 255 patients with pulmonary tuberculosis aged 15 years or older were treated in Cape Town. The number declined annually to 19 174 in 2014, the year after the test roll-out was completed. Table 1 shows the patients' demographic and clinical characteristics in each calendar year: their mean age was 35 years, 57% (57 664/102 007) were male and 48% (47 542/100 021) were HIV-positive. The HIV status was reported in 98% (100 021/102 007) of the patients. The only characteristic that changed substantially over time was the proportion of patients previously treated for tuberculosis, which was lower in later years. By March 2012, 62% (63/101) of tuberculosis clinics had access to the test. There was no difference in patients' demographic or clinical characteristics by either year relative to test roll-out or calendar year. Details of the patients covered by rapid testing in each year and their characteristics are available from the corresponding author on request.
The pattern of microbiological testing in patients with pulmonary tuberculosis changed during the test roll-out: use of this test increased to 88%  Table 3). The principal change underlying the decrease in empirical treatment during the study period was that fewer  among HIV-positive patients, the corresponding figures were 15% (1544/9985) and 1% (84/8823), respectively. In con-trast, the proportion treated despite a negative Xpert® test result did not change substantially over time relative to test roll-out and the proportion treated because of abnormal chest X-ray findings alone decreased slightly among HIVnegative patients but did not change among HIV-positive patients, by both time definitions (details available from the corresponding author on request).
The tuberculosis notification rate in the adult population decreased over the 5-year study period: by 12% among HIV-negative individuals and by 19% among HIV-positive individuals ( Fig. 3 and Fig. 4, respectively). The rate of bacteriologically confirmed tuberculosis increased by 1% and 3% in these two groups, respectively, and the rate of empirical treatment decreased by 56% and 49%, respectively. A slightly different pattern was seen when the data were analysed by year relative to test roll-out: the rate of bacteriologically confirmed tuberculosis increased between year 0 and year 4 by 7% in HIV-negative individuals (Fig. 5) and by 17% in HIVpositive individuals (Fig. 6), and the rate of empirical treatment decreased by 47% and 37%, respectively. These changes stabilized after year 3.
Multivariable logistic regression analysis showed that the odds of empirical tuberculosis treatment were 2.75-fold higher in HIV-positive than HIV-negative patients (adjusted odds ratio: 2.75; 95% confidence interval: 2.55-2.98). Other factors associated with empirical treatment were all patients older than 45 years and female sex in HIV-infected patients. After adjusting for these factors, the odds of empirical treatment decreased with time relative to test roll-out (Table 4). There was no evidence to support a further reduction in odds between years 2, 3 and 4 (P = 0.22). When the analysis was performed separately in HIV-positive and HIV-negative patients, the same risk factors were identified (Table 4). Among HIV-positive patients, every 50-cells/ mm 3 increase in CD4+ cell count at tuberculosis diagnosis was associated with 4% lower odds of empirical treatment. These results were found to be robust in the sensitivity analysis performed using a random effects model (details available from the corresponding author on request).

Discussion
We found that the introduction of the Xpert® test as the first-line diagnostic test for tuberculosis in a large population cohort led to this test becoming the primary method of diagnosis in three quarters of adults treated for drug-sensitive pulmonary tuberculosis. In addition, the rate of bacteriologically confirmed dis-   Rapid diagnostic test roll-out, South Africa Sabine Hermans et al.
ease increased following the introduction of the test and the rate of empirical treatment decreased, resulting in a net decline in the total notification rate for pulmonary tuberculosis. These changes occurred cumulatively with test roll-out and stabilized after 3 years. Few evaluations of the impact of routine Xpert® testing in programmatic settings have been published and most documented difficulties with roll-out and implementation. [24][25][26][27][28][29] Increased proportions of bacteriological confirmation and less empirical treatment were reported in Nepal and India but an increase in the case notification rate was reported only in India. 30,31 Moreover, the four large randomized clinical trials performed to date all reported increased proportions of bacteriological confirmation but only one, performed in Cape Town, 11 found that the number of patients diagnosed with tuberculosis increased. [10][11][12][13] Our data from a programmatic setting in Cape Town are consistent with previous findings: routine use of the rapid test did not lead to an increase in the tuberculosis notification rate but was temporally associated with an increased bacteriological confirmation rate and a decrease in empirical treatment. The net effect was an apparent decline in the total tuberculosis notification rate. We previously reported that notification rates have decreased since 2010 in both HIV-negative and HIV-positive individuals. 32 There are two potential, complementary explanations: (i) the incidence of tuberculosis decreased (assuming access to diagnosis did not change); and (ii) empirical treatment decreased. It was not possible to separate the contributions of these factors using our data. However, the observation that the rate of bacteriologically confirmed tuberculosis remained stable with the increasing use of a more sensitive test suggests that the incidence of tuberculosis may have decreased. A possible underlying mechanism could be greater use of antiretroviral therapy in the HIV-infected population -coverage increased from 0% in 2004 to 63% in 2013 in Cape Town. 32 However, the decline in tuberculosis notification rates we observed was not affected by HIV status, which does not support this explanation. In contrast, the possibility that the empirical treatment rate decreased with the test roll-out is supported by our observation that the decline in this rate HIV: human immunodeficiency virus. a Patients who initiated tuberculosis treatment in the year before the Xpert® MTB/RIF test roll-out date for their clinic were assigned to year 0, those who started treatment in the first year after roll-out for their clinic were assigned to year 1 and so on up to year 4 ( Fig. 1)  Research Rapid diagnostic test roll-out, South Africa Sabine Hermans et al.
slowed gradually during roll-out and then stabilized. This may reflect the time needed to fully implement the new test and, possibly, to apply the new diagnostic algorithm. If the decrease in the tuberculosis notification rate we observed were mainly attributable to a reduction in empirical treatment, we would expect the decline to stabilize within 3 to 4 years of the test being introduced at the last clinic (i.e. by the end of 2017). If the decline continues thereafter, it is probably attributable to another factor, such as declining incidence. 12 The main limitation of our study was its inability to determine whether the decline in empirical treatment represented a decline in false-positive or true-positive diagnoses. The latter could have occurred because clinicians overestimated the negative predictive value of the rapid test. Interestingly, the proportional decline in empirical treatment was smaller among HIV-infected patients, in whom the test is less sensitive. 3 Clinicians may have been reluctant to miss active tuberculosis disease in this vulnerable population, thereby increasing the number of false-positive diagnoses. Our lack of data on presumptive tuberculosis patients precluded an evaluation of whether roll-out of the test led to the identification and treatment of patients who would otherwise not have been treated. Moreover, we were not able to investigate the impact of symptomatology on the likelihood of empirical treatment (no data) or of the time to treatment initiation (incomplete recording of sputum collection dates). The apparent decrease in previously treated patients was probably due to misclassification following abolition of the distinct retreatment regimen.
In conclusion, routine use of the Xpert® test in a setting with a high prevalence of tuberculosis and HIV infection was associated with a halving of the empirical treatment rate. This reduction occurred gradually following the introduction of the test, probably due to the time needed for full implementation of a new diagnostic algorithm. More than a quarter of HIV-infected patients were still treated empirically, which highlights the difficulty of diagnosing tuberculosis in this group. ■  Year relative to rapid molecular diagnostic test roll-out

All patients Patients with bacteriologically confirmed tuberculosis Patients who received empirical treatment
Year 0 Y ear 1 Y ear 2 Y ear 3 Y ear 4 HIV: human immunodeficiency virus. Notes: Patients who initiated tuberculosis treatment in the year before the Xpert® MTB/RIF test (Cepheid, Sunnyvale, United States of America) roll-out date for their clinic were assigned to year 0, those who started treatment in the first year after roll-out for their clinic were assigned to year 1 and so on up to year 4 (Fig. 1). Treatment was empirical when no test gave a positive result or no test was performed.  Year relative to rapid molecular diagnostic test roll-out