Prioritising attributes for tuberculosis preventive treatment regimens: a modelling analysis

Abstract

Background: Recent years have seen important improvements in available preventive treatment regimens for tuberculosis (TB), and research is ongoing to develop these further. To assist with the formulation of target product profiles for future regimens, we examined which regimen properties would be most influential in the epidemiological impact of preventive treatment.

Methods: Following expert consultation, we identified 5 regimen properties relevant to the incidence-reducing impact of a future preventive treatment regimen: regimen duration, efficacy, ease-of-adherence (treatment completion rates in programmatic conditions), forgiveness to non-completion and the barrier to developing rifampicin resistance during treatment. For each regimen property, we elicited expert input for minimally acceptable and optimal (ideal-but-feasible) performance scenarios for future regimens. Using mathematical modelling, we then examined how each regimen property would influence the TB incidence reduction arising from full uptake of future regimens according to current WHO guidelines, in four countries: South Africa, Kenya, India and Brazil.

Results: Of all regimen properties, efficacy is the single most important predictor of epidemiological impact, while ease-of-adherence plays an important secondary role. These results are qualitatively consistent across country settings; sensitivity analyses show that these results are also qualitatively robust to a range of model assumptions, including the mechanism of action of future preventive regimens.

Conclusions: As preventive treatment regimens against TB continue to improve, understanding the key drivers of epidemiological impact can assist in guiding further development. By meeting these key targets, future preventive treatment regimens could play a critical role in global efforts to end TB.

Keywords: Mathematical modelling; Preventive therapy; Tuberculosis.

Fig. 1

Schematic illustration of the model structure. For clarity, the figure concentrates on model compartments relevant to latent TB infection, its progression to active disease, and the effect of preventive treatment on these dynamics; further information on the care cascade for active TB disease (shown in the dotted rectangle) is provided in the supporting information. Each compartment shown here is further stratified by the HIV status. Amongst compartments relating to the natural history of TB, U denotes ‘uninfected’; L^(f)denotes latent infection with ‘fast’ progression (in the first 2 years following infection); L^(s) denotes latent infection with ‘slow’ progression; and I denotes active, infectious TB. The modelled action of preventive treatment is as follows: S denotes individuals who are bacteriologically cured of infection as a result of the regimen; the parameter c governs the proportion cured in this way. Q denotes individuals with non-curative, post-regimen protection; the parameter e denotes the strength of non-curative protection, while g denotes its post-regimen durability. P1 and P2 represent individuals who are, respectively, undergoing the first and second halves of the regimen. Reinfection is possible for stages P1, P2, S, Q, at a reduced rate compared to R, given an assumed level of protection from preventive therapy. Reinfection is not shown for clarity, but is modelled by transitions from P1, P2, Q and R to their respective f states (e.g., P1^(f)), and from S to L_f. We assume that ‘forgiveness to non-completion’ (one of the regimen properties listed in Table 2) applies only to the latter, with the parameter f being the proportion interrupting treatment who nonetheless have the same outcomes as those completing treatment. R denotes individuals who have reverted to their pre-regimen state of TB infection, following decay of any post-regimen protection. d denotes the per-capita hazard of regimen interruption and thus governs ease-of-adherence, b directly models the drug-resistance barrier and m governs the regimen duration. These and remaining model parameters are as listed in Additional file 2: Table S2

Fig. 2

The influence of different regimen properties on potential incidence reductions from a future preventive treatment regimen. Shown are partial rank correlation coefficients (PRCCs) between each regimen property listed in Table 2, and the percent cases averted between 2020 and 2035, of a regimen that is rolled out to cover all PLHIV on ART, as well as all household contacts of notified cases. Larger bars indicate regimen properties having greater influence on incidence reductions; error bars show 95% uncertainty intervals, estimated by bootstrapping. See Fig. S9 for scatter plots underlying these correlations

Fig. 3

Contribution of regimen properties to epidemiological impact. While Fig. 2 shows regimen properties in order of decreasing influence on incidence reductions, this figure shows the quantitative effect of each on impact. In each panel, the leftmost bar shows the impact (cumulative cases averted) of a PT regimen with properties fulfilling only minimal criteria in Table 2. The remaining bars show the impact of successively optimising each single regimen property in turn, starting with the most influential properties shown in Fig. 2. For clarity, only the top four most influential properties are shown. The horizontal dashed line shows the impact arising from a fully optimised regimen, i.e., with all 5 properties assuming optimal values in Table 2). Error bars, and the gray shaded area, show 95% Bayesian credible intervals on respective estimates

Fig. 4

The influence of ‘mechanistic’ regimen parameters on regimen efficacy. As described in the main text, we define ‘efficacy’ as the incidence reductions that would result under trial conditions, over 2 years of follow-up. However, efficacy in the model accommodates a range of scenarios for the relative role of different, underlying mechanisms of protection. Shown are estimates for the strength of association between each mechanistic parameter, and efficacy (incidence reductions) that would be measured under trial conditions, upon 2-year follow-up. As in Fig. 2, larger bars indicate those parameters that are more influential for efficacy, and error bars show 95% uncertainty intervals, estimated through bootstrapping. See Additional file 5: Fig. S10 for further analysis on how parameters interact to yield efficacy