Huifang You1, Jyothi Gurajala1, Ulrika simonsson1
1Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden, 2Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden, 3Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden
Introduction: In tuberculosis (TB) drug development, Phase 2a trials assess early bactericidal activity over two weeks using the time-to-positivity (TTP) biomarker (1,2), followed by Phase 2b trials, which evaluate an 8-week treatment period with culture conversion as the endpoint (3). Phase 2a data can evaluate treatment efficacy for individual drugs or different regimens and thereby support the design of Phase 2b trials. Compared to culture conversion at 2 months, TTP provides a more detailed quantification of bacterial kill over time and can potentially be more informative than just evaluating the efficacy at 2 months. However, it is not completely understood how Phase 2a TTP data in TB drug development can be used to guide Phase 2b study design to optimize trial efficiency and decision-making. In this study, we aimed to predict Phase 2b TTP over time using 2-week Phase 2a TTP data using machine learning (ML). Additionally, we investigated how different Phase 2a study durations informed the predictive performance of longitudinal TTP biomarker information in Phase 2b. Methods: Time-to-positivity (TTP) data were obtained from a randomized, double-blind, placebo-controlled Phase 3 trial evaluating the noninferiority of two moxifloxacin-containing regimens compared to a control regimen for uncomplicated, smear-positive pulmonary tuberculosis (4). To ensure the stationarity of the dataset, a logarithmic transformation was applied, followed by differencing. The stationary dataset was then used for time-series analysis under two different training and forecasting scenarios: (a) training on data up to 2 weeks with forecasting from 2 weeks to 2 months and (b) training on data up to 4 weeks for forecasting from 4 weeks to 2 months. The following features were used as predictors for TTP: height, weight, regimen, presence of cavitation, smoking status, and time since the start of treatment. Four machine learning models were applied using Python (version 3.9.1) to predict TTP over time up to 2 months: AutoRegressive Integrated Moving Average (ARIMA), AutoRegressive Integrated Moving Average Exogenous Variables (ARIMAX), extreme gradient boosting (Xgboost), and Random Forest (RF). Hyperparameter tuning methods were applied as appropriate for each model. Model performance was assessed using root mean squared error (RMSE) and mean absolute error (MAE), and visual predictive checks (VPCs) with observed mean TTP compared to 90% confidence interval from the final model which guiding model selection. Results: The TTP data were well captured by using time-series prediction models, as confirmed by VPCs. The ARIMA model was identified as the best-performing model for a scenario with 2 weeks of training data and forecasting TTP from 2 weeks to 2 months, achieving an RMSE of 11.8 and an MAE of 8.9. For the scenario with training on data for up to 4 weeks and forecasting TTP from 4 weeks to 2 months, the ARIMA model showed the best performance with an RMSE of 13.1 and an MAE of 9.6. For both scenarios, the good fit by the ARIMA model was indicated by the estimated 90% confidence interval compared to the observed mean TTP over time. Additionally, ARIMA exhibited similar performance across both scenarios, implying that adding 4-week biomarker data did not improve the TTP predictions compared to using only 2-week biomarker data. Conclusions: Machine learning provides opportunities for forecasting longitudinal biomarker data within a Phase 2b study for tuberculosis drug development using only two-week Phase 2a information. Among the machine learning models tested, ARIMA consistently performed the best, resulting in good accuracy for the two different scenarios. These findings can support the translation of Phase 2a results into traditional Phase 2b trial design but also adaptive trial designs by continuously updating TTP predictions over time for different treatment regimens.
1. Diacon AH, Maritz JS, Venter A, van Helden PD, Dawson R, Donald PR. Time to liquid culture positivity can substitute for colony counting on agar plates in early bactericidal activity studies of antituberculosis agents. Clinical Microbiology and Infection. 2012 Jul 1;18(7):711–7. 2. Mockeliunas L, Faraj A, Van Wijk RC, Upton CM, Van Den Hoogen G, Diacon AH, et al. Standards for model-based early bactericidal activity analysis and sample size determination in tuberculosis drug development. Front Pharmacol. 2023 Apr 13;14:1150243. 3. Svensson EM, Svensson RJ, Te Brake LHM, Boeree MJ, Heinrich N, Konsten S, et al. The Potential for Treatment Shortening With Higher Rifampicin Doses: Relating Drug Exposure to Treatment Response in Patients With Pulmonary Tuberculosis. Clin Infect Dis. 2018 Jun 18;67(1):34–41. 4. Gillespie SH, Crook AM, McHugh TD, Mendel CM, Meredith SK, Murray SR, et al. Four-Month Moxifloxacin-Based Regimens for Drug-Sensitive Tuberculosis. N Engl J Med. 2014 Oct 23;371(17):1577–87.
Reference: PAGE 33 (2025) Abstr 11763 [www.page-meeting.org/?abstract=11763]
Poster: Methodology – AI/Machine Learning