Applying a Pharmacometrics-Enabled Machine Learning Analysis to Predict 2-Month Culture Conversion Using Phase 2a Data in a Tuberculosis Clinical Trial.

Phase 2a trials in tuberculosis patients traditionally assess early bactericidal activity over two weeks, often using the time-to-positivity biomarker, followed by a phase 2b study typically lasting 8-week with time-to-event of culture conversion as the endpoint. This study investigated different machine learning models to predict the time-to-event of 2-month culture conversion in the REMoxTB trial with phase 2a time-to-positivity biomarker data and the impact of different phase 2a study lengths. Time-to-positivity at baseline and up to 14 days or 28 days after two moxifloxacin-containing regimens and one control regimen were analyzed using nonlinear mixed-effects modeling. The final models were used to predict individual baseline time-to-positivity and time-to-positivity differences between 0 and 14 days or 0 and 28 days. The individual predictions served as features in the following machine learning analysis. Statistical metrics and Kaplan-Meier plots informed model selection. Bi-exponential and exponential decay models described the 14-day and 28-day time-to-positivity data, respectively. In the machine learning analysis, baseline time-to-positivity and/or time-to-positivity differences were ranked as the most important features in all models. Statistical metrics and Kaplan-Meier plots indicated a good fit to culture conversion at 8 weeks using 4-week phase 2a information with a C-support vector classification model. Four-week time-to-positivity phase 2a data provided more information compared to the 2-week time-to-positivity data for the prediction of culture conversion. The workflow demonstrated the potential of machine learning to predict the time-to-event of phase 2b culture conversion up to 8 weeks using time-to-positivity phase 2a biomarker information in a clinical trial assessed with pharmacometric analysis.