Target prioritization and strategy selection for active case-finding of pulmonary tuberculosis: a tool to support country-level project planning

A simple deterministic model was designed to calculate the estimated pulmonary TB cases diagnosed and the associated direct costs for different diagnostic algorithms (Figure 1), with the parameters and assumptions listed in Table 1. The screening of TB involves multiple steps of several testing methods, usually consisting of one or two steps of screening tests followed by a confirmatory diagnosis with bacteriological tests.

The primary output of the model was the diagnostic cost per TB case detected, defined as the total diagnostic cost divided by the total number of TB cases diagnosed. The total diagnostic cost was derived from a simple step-by-step calculation of costs. The total number of TB cases diagnosed was calculated from the number of subjects screened, the TB prevalence among them, the proportion of prevalent TB cases diagnosed with a given set of diagnostic procedures (proportional yield) and the dropout rate. In the model, dropouts were defined as individuals who entered into a screening/diagnostic step but did not complete the whole step through to receiving the test result. In designing the model, TB prevalence was defined as the prevalence of bacteriologically-confirmed pulmonary TB.

Six diagnostic algorithms were predefined to examine their cost-effectiveness when applied to different target populations (Table 2). As a rapid guide, the strategies were constructed from Strategy 1 to Strategy 5 in an incremental manner from a simple, conservative approach to extensive approaches. Strategy 1, symptom screening followed by sputum microscopy, is similar to routine programme settings. Strategy 2 adds X-ray for diagnosis of smear-negative cases. Strategy 3 employs radiography screening for the entire target population, and Strategy 4 is a standard approach recommended by WHO for prevalence surveys [12]. Strategy 5 uses Xpert as the diagnostic test with the screening procedure of Strategy 4. Strategy 6 is the only strategy with two-step screening. The symptomatic subjects identified by the first screening undergo X-ray screening to reduce the number of subjects for the diagnostic test (Xpert). This algorithm reduces the number of Xpert tests at the cost of losing non-symptomatic TB patients at the first screening step.

All six diagnostic algorithms employ one of the following three screening approaches to identify subjects proceeding to the diagnostic tests: (i) TB symptomatic screening, i.e., typically, inquiring about the presence of prolonged cough for more than two weeks and/or haemoptysis; (ii) a combination of TB symptomatic screening and chest radiography, wherein the model assumes that individuals with TB symptoms and/or abnormal chest radiography would be ruled in; and (iii) a sequential application of symptom screening and chest radiography which identifies individuals with both TB symptoms and any abnormal chest radiography.

The first approach is bound to miss a substantial portion of TB patients (45%–61%) [4‐6] who do not fulfil TB symptomatic criteria. The second approach is one of the most inclusive screening methods and has been recommended by WHO for TB prevalence surveys [12]. The basic idea of the third approach is to reduce the number of subjects who undergo an expensive diagnostic test. In our model, it was used only in Strategy 6, aiming to reduce the cost of tests using Xpert [11].

To calculate the number of subjects proceeding to the next step, the estimated prevalence of abnormality among the screened subjects was required. The information from prevalence surveys [4, 6, 12, 13] was used to construct simple linear regression models in order to predict the prevalence of abnormalities as a function of TB prevalence (Table 1).

A proportional yield, defined as the maximum proportion of prevalent TB cases diagnosed by a given diagnostic algorithm, was required to calculate the number of TB cases diagnosed. In other words, the proportional yield can be interpreted as the sensitivity of diagnostic methods in combination. The assumptions needed for estimating the proportional yield were distributions of prevalent TB cases by symptom and smear status as shown in Figure 2(a). These numbers can be derived from empirical data from prevalence surveys but, as shown in Figures 2(b) and 2(c), the survey findings were not fully consistent [4, 6]. While more data are needed from other surveys, the yield can be theoretically calculated from two assumed parameters: the sensitivities of TB symptom screening and sputum smear microscopy. Figure 2(d) shows proportional yields that were calculated by assuming that 40% of TB cases were symptomatic and 60% were smear-positive. Based on a chosen set of the proportional yield, the number of TB cases diagnosed can be calculated by combining the segments in Table 2. Dropout rates defined in Table 1 were used to discount the eventual number of cases who could be diagnosed.

The model intends to produce cost-effectiveness parameters for multiple risk groups at once so that the results can be compared to judge the relative importance of different risk groups. Estimated TB prevalence can be defined by two methods depending on the type of group. TB prevalence of some groups may be better defined as a prevalence ratio to the baseline prevalence (e.g. smokers) while other groups may have direct estimates of TB prevalence if relevant data is available (e.g. TB contacts, prisoners). For the model analysis, we arbitrarily defined the four TB risk groups by a prevalence ratio (PR=1.0, 1.5, 2.0 and 3.0) and two high-risk groups by prevalence estimates of 4000 and 6000 cases per 100,000 population (Table 1). A detailed discussion on the estimates of TB risk is not within the scope of this study. Nevertheless, literature is available on the estimates of differential risk of TB among various groups to make inference from the model outputs to a real-life scenario [14‐18]. For the online tool, users are expected to set estimated prevalence parameters according to the local context and the model provides context-specific outputs to facilitate local level project planning.

Under a given baseline of TB prevalence among the general population, the model estimates a diagnostic cost per case detected for each of the defined high-risk populations. Cost per case detected has been used in the global TB control community through two international initiatives: Fund for Innovative DOTS Expansion Through Local Initiatives to Stop TB (FIDELIS) by the International Union Against Tuberculosis and Lung Disease launched in 2004 [19], and TB REACH by the Stop TB Partnership launched in 2010 (http://​www.​stoptb.​org/​global/​awards/​tbreach). TB REACH generally recommends formulating a project within the overall budget of USD 350 per case detected and successfully treated. Since the model only calculates the direct diagnostic cost, not the costs associated with operation, logistics and treatment, USD 200 per case detected was arbitrarily used as a benchmark. Additional outputs include the Number Needed to Screen (NNS) and incremental yields in relation to the total costs incurred. Sensitivity analysis was conducted for the selected parameters for which values were less certain.

The model was initially developed in Microsoft Excel, and summary outputs were utilized for formulating ACF projects in selected countries in the Western Pacific Region. After some upgrading, the tool was coded in the statistical package R (CRAN: the Comprehensive R Archive Network at http://​cran.​r-project.​org) to conduct statistical analysis and simulations.

As an attempt to provide the model on a user-friendly and interactive platform, the tool was rewritten in JavaScript to be available on the Internet at http://​www.​innovationsinpub​lichealth.​org.

Figure 3(a) shows the yield of TB cases for each of the six diagnostic strategies against overall diagnostic cost under a hypothetical setting of 10,000 population with 1.0% of TB prevalence (1000 per 100,000 population). In general, the number of TB cases diagnosed increased as the diagnostic strategy moved from conservative to extensive (strategies 1 to 5) with an additional diagnostic cost at each step. Cost per case detected also linearly increased, except for Strategy 6, as shown in Figure 3(b).

Under the hypothetical setting, Strategy 1 diagnosed only 23.5 patients out of 100 prevalent TB cases in the population with a diagnostic cost of USD 1571. In terms of cost-effectiveness, this translates to USD 67 per case detected. With Strategy 2, a total of 36.1 patients (an additional 12.6 patients) were diagnosed with an overall diagnostic cost of USD 3530. Strategies 3 and 4 substantially increased the number of TB cases (90.2 and 91.7 patients, respectively), but the cost was substantially higher (more than tenfold, USD 32 535 and USD 43 655, respectively). Cost per case detected became USD 361 and USD 476 which is several times higher (i.e. less cost-effective) than the first two strategies. The increased cost for Strategy 3 was primarily due to X-ray examinations for the entire targeted population, and the difference between the two strategies (3 and 4) was the use of culture as a diagnostic test. The model assumed that Xpert yielded the same level as solid culture, but a relatively high dropout rate (15%) was set for culture considering the long turnaround time. As expected, Strategy 6 yielded a relatively small proportion of TB cases because the strategy applied the most restrictive approach in screening. However, it is important to note that this strategy seemed reasonably cost-effective with USD 155 per case detected, despite the higher cost of using of Xpert.

Figure 4(a) summarizes the cost-effectiveness of diagnostic strategies for different risk groups. Each of the six plotted lines represents a different risk group with PR of 1.0 (general population), 1.5 (e.g. poor urban communities), 2.0 (e.g. malnourished), 3.0 (e.g. diabetics), 8.0 (e.g. TB contacts) and 12.0 (e.g. prisoners or people living with HIV). The baseline TB prevalence was set at 0.5%, which is approximately the prevalence level for Bangladesh, the Democratic Republic of the Congo, Mozambique, Myanmar and the Philippines, and is lower than the prevalence for Cambodia and South Africa [1].

As observed in Figure 3, the general trend of decreasing cost-effectiveness with upgraded diagnostic strategies persisted for all risk populations. However, the cost per case detected was substantially reduced in a population with a very high prevalence such as TB contacts, prisoners or people living with HIV. In these very high-risk populations, the most extensive diagnostic strategies (such as the strategies 4 and 5) were still cost-effective (i.e. all strategies fell within a diagnostic cost per case detected of less than USD 200).

Another important finding was that conservative approaches which were similar to routine programme setups (strategies 1 and 2) were the most cost-effective strategies even though the yield was low. These strategies were cost-effective for almost all high-risk populations in TB high-burden countries.

The cost per case detected increased abruptly between Strategy 2 and Strategy 3 due to the application of X-rays for all subjects. With the given set of assumptions and using USD 200–250 as a cost-effectiveness target, TB prevalence around 1.5% was identified as the decision-making-point whether to employ X-ray screening for all. However, it requires context-specific analysis because the cost-effectiveness of Strategy 3 appears sensitive to the unit cost of X-ray which varies among different countries and localities. This issue was further examined in sensitivity analyses.

The Number Needed to Screen, defined as the total number of subjects screened divided by the number of TB cases detected, is presented in Figure 4(b). Obviously, the higher the TB prevalence, the lower the NNS. However, even within the same risk group, the NNS significantly differed depending on the diagnostic strategy. Under the assumptions for Figure 4, NNS went up to several hundred for the low-yield strategies (1 or 2), but the diagnostic cost was actually very low. Conversely, the NNS was minimized by applying extensive diagnostic strategies that yielded more cases but were costly. Due to this inverse relationship between the NNS and the diagnostic cost, a low NNS would not necessarily guarantee the feasibility and cost-effectiveness of ACF. The reality seems the other way round - inexpensive strategies (1 and 2) were almost always cost-effective even though the NNS was very high.

The sensitivity of the cost per case detected was examined under varied parameter conditions. Figure 5(a) presents the result when the unit cost of X-ray uniformly varied between USD 2 and 6. As expected, the cost per case detected varied significantly for strategies 3, 4 and 5. Importantly, the output was more robust under the prevalence of 4.0% than under 1.0%.

Another less certain set of assumptions was the proportional yield. The uncertainty bound was obtained by uniformly, simultaneously and independently varying the proportion of symptomatic patients between 30% and 50% and the proportion of smear-positive patients between 50% and 70% (Table 1). As a result, Figure 5(b) shows that the model output was reasonably robust against the proportional yield assumptions, which minimized the issue of uncertainty due to insufficient empirical data.