Analysis of time-to-positivity data in tuberculosis treatment studies: Identifying a new limit of quantification

Abstract

The BACTEC Mycobacteria Growth Indicator Tube (MGIT) machine is the standard globally for detecting viable mycobacteria in patients' sputum. Samples are observed for no longer than 42 days, at which point the sample is declared "negative" for tuberculosis (TB). This time to detection of bacterial growth, referred to as time-to-positivity (TTP), is increasingly of interest not solely as a diagnostic tool, but as a continuous biomarker wherein change in TTP over time can be used for comparing the bactericidal activity of different TB treatments. However, as a continuous measure, there are oddities in the distribution of TTP values observed, particularly at higher values. We explored whether there is evidence to suggest setting an upper limit of quantification (ULOQ_M) lower than the diagnostic limit of detection (LOD) using data from several TB-PACTS randomized clinical trials and PanACEA MAMS-TB. Across all trials, less than 7.1% of all weekly samples returned TTP measurements between 25 and 42 days. Further, the relative absolute prediction error (%) was highest in this range. When modeling with ULOQ_Ms of 25 and 30 days, the precision in estimation improved for 23 of 25 regimen-level slopes as compared to models using the diagnostic LOD while also improving the discrimination between regimens based on Bayesian posteriors. While TTP measurements between 25 days and the diagnostic LOD may be important for diagnostic purposes, TTP values in this range may not contribute meaningfully to its use as a quantitative measure, particularly when assessing treatment response, and may lead to under-powered clinical trials.

Keywords: MGIT; clinical trials; early bactericidal activity; time-to-positivity; tuberculosis.

Fig 1.. Observed time-to-positivity trajectories in REMoxTB.

Any observations at or above the diagnostic limit of detection (42 days) are recorded as 42 days. A: Regimen-level trends in TTP (lines) and estimated STAND, PaMZard errors (ribbons) as fit by smoothing splines. B: Individual TTP trajectories (light gray) and regimen-level smoothing spline (black).

Fig 2.

The flow of individuals’ weekly time-to-positivity (TTP) samples from measurements of ≤ 25 days, between 25 and 30 days, between 30 and 42 days, and above the diagnostic LOD (≥ 42 days) for A) REMox-TB, B) PanACEA MAMS-TB, C) NC-002 (PaMZ), D) NC-005 (BPaMZ), E) NC-006 (STAND, PaMZ), F) Study 29, and G) Study 29X.

Fig 3.

A) Replicate TTP observations from the NC-002 (PaMZ) study plotted against each other on the log₁₀ scale. A red line indicates where perfect replication would lie. Points are marked in black (•) if below the diagnostic LOD and yellow (•) if above the diagnostic LOD. B) Absolute prediction error (%) when using one observation from each replicate pair (TTP₁) to predict the second observation from each replicate pair (TTP₂), using 5-fold cross-validation to train and test a simple linear prediction model. Data used for prediction is restricted to only those observations below the diagnostic limit of detection. The solid line denotes the average absolute prediction error (%) within deciles of TTP₁ observations. The dashed line denotes a LOESS fit. Some predictions had an absolute prediction error (%) greater than 100% and are marked by triangles (▲).

Fig 4.

Among regimens with improved bactericidal activity over HRZE, the posterior distributions for the relative comparison of a regimen’s slope $\left({\gamma }_{1j}\right)$ against the estimated slope on HRZE (${\gamma }_{1,\mathrm{H}\mathrm{R}\mathrm{Z}\mathrm{E}}$), where a value of 1 indicates equal slopes and values > 1 suggest the regimen has greater bactericidal activity than HRZE. The estimated “confidence” that a regimen has any improvement in bactericidal activity over HRZE $\left(\mathrm{P}\mathrm{r}\left({\gamma }_{1j}/{\gamma }_{1,\mathrm{H}\mathrm{R}\mathrm{Z}\mathrm{E}}>1\right)\right)$ as well as the “confidence” that a regimen has more than 10% improvement in bactericidal activity over HRZE $\left(\mathrm{P}\mathrm{r}\left({\gamma }_{1j}/{\gamma }_{1,\mathrm{H}\mathrm{R}\mathrm{Z}\mathrm{E}}>1.1\right)\right)$ is indicated for each regimen at each ULOQ_M.