Multiple clinical trials over the last 40 years attempting to shorten therapy for pulmonary, drug-sensitive tuberculosis from six to four months have failed ^{1– 4} . Despite this, the various four-month treatment arms consistently cured about 80–85% of patients. The premise of the ongoing PredictTB clinical trial is that this sub-population of lower risk patients who are cured at four months can be identified prospectively ⁵ . Two subsequent secondary analyses of the three fluoroquinolone treatment shortening trials ^{2– 4} support this premise by finding that a subset of lower risk participants, those without cavity on baseline chest-x-ray, acid-fast bacilli (AFB) smear less than 2+ at baseline, or AFB smear negative at month 2 of treatment were associated with successful four-month treatment outcomes ^{6, 7} . Another trial prospectively attempted a similar risk stratification by shortening treatment for those without cavity on baseline chest x-ray, a measure of baseline risk, and those with negative sputum cultures at two months of treatment, a measure of treatment response ⁸ . Although this trial also failed, the treatment success rate in the four-month arm achieved 93%, higher than other four-month treatment trials that did not risk stratify participants. The PredictTB trial was developed on the hypothesis that more precise methods of evaluating baseline risk and treatment response will successfully identify lower risk participants cured with four months of treatment.

The value of early chest x-ray changes to predict TB treatment outcomes has been recognized for over 60 years ⁹ . Cavities on baseline chest x-ray ^{10– 12} and residual cavity at cure ¹³ have been associated with poor treatment outcomes. Computed tomography scans are more sensitive than x-ray. Further, changes on 2-deoxy-2-[ ¹⁸F]fluoro-D-glucose (FDG)-positron emission tomography/computed tomography (PET/CT) scans have been associated with treatment outcomes in nonhuman primates ^{14, 15} . Results in patients have been limited to small numbers due to the difficulty of collecting PET/CT imaging in this setting. However, a study of patients with pulmonary, multidrug-resistant tuberculosis (MDR-TB) showed a relationship with long-term treatment outcomes ¹⁶ . In drug sensitive TB (DS-TB), distinct response patterns on PET/CT scan from baseline to month 6 on treatment correlated with treatment outcomes ¹⁷ .

Month 2 sputum culture conversion has traditionally been considered the best biomarker of treatment outcome available ¹⁸ despite a meta-analysis demonstrating poor sensitivity and specificity in predicting outcomes ¹⁹ . This has been confirmed by additional analyses of the REMoxTB treatment shortening trial ² , where month 2 culture conversion status poorly predicted final outcomes ²⁰ . Time-delays associated with culture conversion tests can be as long as 6-8 weeks, resulting in delayed assessments of the patient’s bacillary status. Immediately available test results are likely to predict outcomes better than delayed results because the results reflect the patient’s current status rather than a historical status from 1-2 months ago and thus poor results can be acted upon immediately. The value of contemporaneous results has been shown with CD4 cell counts in HIV ²¹ , as well as in TB where a month 6 culture conversion status predicted final outcomes significantly better than month 2 culture conversion status ²² . A point-of-care test that measures TB bacterial load has the potential to predict outcomes better than a delayed culture conversion result.

The aim of this analysis is to describe how the early treatment completion criteria used in the ongoing PredictTB clinical trial (NIH IRB #16IN133; NCT02821832) were developed. In this analysis, we assess the relationship between measures of baseline risk and treatment response with poor treatment outcomes in a cohort of pulmonary DS-TB patients treated programmatically in South Africa. Baseline risk is assessed using quantitative markers from FDG-PET/CT scans at baseline. Early treatment response is assessed by changes in these markers on FDG-PET/CT scan at week 4 of treatment. Adherence is also assessed, as is a quantitative measure of residual bacterial load using sputum Xpert MTB/RIF cycle threshold at week 16.

Among the 92 patients that we analyzed from the Catalysis cohort, 73 were cured (asymptomatic two years after the end of treatment), eight failed treatment, and 11 programmatically restarted TB treatment during follow-up. For the PET/CT imaging analysis, the baseline and week 4 PET/CT scans were each read by a single reader, with overall summary statistics presented in Table 1. At baseline, cured patients were significantly different from treatment failure patients in CT cavity air volume, with CT hard volume and PET TLG differences being borderline significant (P=0.059 for both). At week 4, only the difference in total cavity air remained significantly different. In contrast, the results for patients who were cures and retreatments were not significantly different from each other in any parameter at baseline or at week 4, making it very difficult to differentiate these two cohorts using these parameters. The comparison of treatment failures with retreatments was similar to that of treatment failures with cures. We therefore developed our criteria based primarily on differences between the cured and failure cohorts.

To identify specific thresholds that predicted cure vs treatment failure, ROC curves were drawn for each variable ( Figure 1). When using the optimal ROC thresholds of all baseline and week 4 PET/CT criteria into a single algorithm, the combined criteria predicted cure with 100% sensitivity, capturing all eight failures and 11 retreatments as high risk. However, specificity was very poor at 19.2% and only 14/81 (17.3%) of subjects with both baseline and week 4 PET/CT scans were classified as low risk ( Table 2). This is well below our target of 50%, classifying too many as false positive high risk, resulting in an algorithm that is neither practical nor scalable. We therefore adjusted the thresholds to be more specific at the cost of sensitivity to approach our 50% target.

Because prior clinical trials data already demonstrated that baseline cavity was a risk factor for poor outcomes ^{8, 10} , and because cavity size was the strongest predictor of poor outcome in our ROC curves (cavity air area-under-the-curve (AUC) > CT hard volume and PET TLG AUCs), we built the algorithm around this parameter first. We defined cavity as largest single cavity size rather than total cavity size because we posited that a single large cavity induced higher risk and may take longer to heal than multiple smaller cavities (i.e. one 30 mL cavity had a higher risk for poor outcome than two 15 mL cavities). In examining the baseline cavity size threshold, a 10.5 mL threshold captured 6/8 (75%) of treatment failures as high risk but only 46/73 (63.0%) cures as low risk. By increasing the threshold to 30 mL, the algorithm would miss one additional failure (now only 5/8 [62.5%] as high risk; increasing beyond 30 mL would lose more than one additional failure) but specificity would increase to 86.3% (63/73 cures now classified as low risk; Table 3). Overall, using 30 mL instead of 10.5 mL as the baseline cavity threshold increases the proportion of cured (N=73) and failure (N=8) patients defined as low risk from 48/81 (59.3%) to 66/81 (81.5%). Changing the week 4 cavity volume reduction threshold from 64% to 20% results in a similar sensitivity/specificity tradeoff. After applying both baseline and week 4 cavity change thresholds, only 29/81 (35.8%) would have been classified as low risk (before adding any other imaging criteria) using the 64% threshold but 58/81 (71.6%) were low risk with the 20% cavity reduction threshold ( Table 3). In contrast to prior data on the risk from baseline cavities, quantitation of CT disease volumes and PET TLG has not previously been validated. Therefore, weighing cavity size to threshold about half (28.4% of total) of the target 50% as high risk seemed appropriate, allowing the remaining criteria (CT hard volume, PET TLG, Xpert cycle threshold, and adherence) to threshold the other half.

Similar to the cavity size thresholds, the cutoffs for hard CT volume and PET TLG were adjusted by decreasing sensitivity but increasing specificity from the optimal ROC parameters ( Figure 1, Table 2) to stratify about 50% of the total cohort as low risk, aiming to capture as many treatment failure and retreatment patients as possible within the 50%. Figure 2 demonstrates the patient stratifications when baseline CT hard volume <200 mL and TLG <1500 units were used as low risk criteria at baseline. When combined with largest cavity air <30 mL, 60/81 (74.1%) patients were classified as low risk at baseline ( Table 4). When applied to the entire Catalysis cohort, including the retreatment patients, 6/8 (75%) failures but only 1/11 retreatments were classified as high risk at baseline, suggesting that treatment failure patients may be more correlated with severity of baseline disease compared to retreatment TB patients. Week 4 change criteria for CT hard volume and PET TLG were similarly adjusted and allowed for slight increases to account for potential paradoxical treatment responses that were ultimately still favorable. The final week 4 criteria selected allowed up to a 10% increase in hard volume and 30% increase in TLG at week 4 to remain low risk, resulting in 46/81 (56.8%) classified as low risk after applying both baseline and week 4 PET/CT criteria ( Table 4). The week 4 criteria captured only one additional failure but two additional retreatments, suggesting that retreatment TB patients may be more correlated with poor treatment response at week 4 rather than severity of disease at baseline.

Finally, we included a measure of residual TB bacterial load in sputum in the early treatment completion criteria based on an analysis that Xpert cycle threshold around 30 at weeks 8 and 24 correlated with culture negativity and patient treatment outcomes ²⁹ . We incorporated this measure at week 16 as a safety mechanism to ensure that participants with higher sputum bacterial load (cycle threshold <30) did not stop treatment early. Combining the baseline and week 4 criteria stratified 41/81 (50.6%) of all cured and failure patients as low risk ( Table 4). When applied to the entire Catalysis cohort, including retreatment patients, 1/8 (12.5%) failures and 6/11 (54.5%) retreatments were captured as low risk. The PredictTB early treatment completion criteria when the trial started is shown in Table 5a.

The initial criteria were established as described above, acknowledging that early changes may be needed once the trial began if the actual proportion of PredictTB study participants stratified to the low- and high-risk arms were not close to the 50:50 target. Indeed, after about nine months of enrollment, only 23.4% of participants to reach week 16 were stratified as low risk (Arms B and C), with the remainder stratified to the high-risk arm (Arm A). This was less than half of the estimated 50% we expected to be low risk, which had major implications for the cost and duration of the study (total sample size, study duration, and cost would need to be increased to achieve required sample size in Arms B and C) as well as the scientific relevance of the study (if trial successful, it would only apply to less than 25% of TB patients, diminishing relevance). After discussion with our study Data and Safety Monitoring Board (DSMB), we revised our study early treatment completion criteria.

Previously conducted treatment shortening studies for DS-TB suggested that approximately 80-85% of patients are cured with four months of treatment ^{1– 4} . Shortening treatment only in lower risk participants who had no cavity on baseline chest x-ray and had sputum culture converted to negative by two months of treatment resulted in the four-month treatment success proportion increasing to 93% in one trial but this was still significantly worse than six months of treatment ⁸ . The PredictTB trial tests an alternate risk stratification criteria based on FDG-PET/CT disease burden at baseline, the change in PET/CT disease burden at week 4 of treatment, and a marker of residual bacterial load and adherence dose count at the end of treatment, hypothesizing that this combination will identify patients with tuberculosis who are cured with four months of standard treatment ⁵ . Risk signatures based on transcriptomics have recently been shown to correlate with treatment outcomes ^{33, 34} .

The development of the PredictTB early treatment completion criteria was based on a cohort of 92 DS-TB patients programmatically treated in Cape Town, South Africa on whom we had PET/CT scans at baseline and week 4 of treatment, Xpert cycle threshold data, and programmatic treatment outcomes (Catalysis cohort). Because these patients were treated programmatically, treatment was not directly observed and we were thus unable to determine the proportion of treatment failures due to poor adherence or differentiate true relapsed disease patients from those re-infected. The lack of these data confounded our attempts to develop early treatment completion criteria that captured treatment failure and true relapse patients with meaningful sensitivity and specificity. We resorted to developing criteria that stratified about 50% of patients as high risk, trusting that the most severely diseased patients at baseline, those with poor treatment responses at week 4, and those under the Xpert cycle threshold cutoff at week 16 were captured as higher risk and therefore not eligible for treatment shortening.

A major limitation of developing our algorithm was the lack of sufficient relapse data to validate our early treatment completion criteria. This limitation is challenging to overcome, given the small numbers of available patient data anywhere with microbiological, strain-type confirmed relapses and the even smaller numbers of these with FDG-PET/CT scan data. We acknowledge the risk of overfitting our data (i.e., producing a risk model that may not be generalizable because it was fit only to the data on which it was developed) and in fact, after the study started, it became clear that our criteria were too conservative, stratifying >75% of participants as high risk and therefore not eligible for treatment shortening. Without immediate correction, we would likely run out of funding before the end of the trial due to the increased total sample size needed to achieve the required lower risk cohort sample size. Even worse, we would end up with a trial result that was applicable only to the 20–25% of patients stratified to the lower risk arms and therefore not relevant to the majority TB patients. Our amended early treatment completion criteria, however, have been stratifying participants at roughly 50:50 to the high- vs. low-risk arms.

The PredictTB trial early treatment completion criteria were developed to identify those with the most severe disease at baseline (potentially at higher risk for treatment failure) and with a poor week 4 treatment response (potentially at higher risk of relapse), along with a marker of residual bacterial load and an adherence dose count at treatment completion. These criteria are currently stratifying about 50% of patients to the higher risk arm and 50% of patients to the two lower risk arms, which is the target goal. Whether or not this will successfully identify a lower risk cohort that can be successfully cured with four months of standard therapy awaits the results of the trial, expected in 2022.