Maxwell T. Chirehwa (1), Helen McIlleron (1), Lubbe Wiesner (1), Dissou Affolabi (2), Oumou Bah Sow (3), Paolo Denti (1), and Corinne Merle (4,5) on behalf of RAFA team
(1) Division of Clinical Pharmacology, Department of Medicine, University of Cape Town, South Africa (2) National Hospital for Tuberculosis and Pulmonary Diseases, Cotonou, Benin (3) Service de Pneumopthysiologie, Hopital Ignace Deen, Conakry, Guinea (4) London School of Hygiene & Tropical Medicine, London, UK (5) UNICEF/UNDP/World Bank/WHO Special Programme on Research and Training in Tropical Disease (TDR), Geneva, Switzerland
Objectives: The long duration of therapy remains a challenge for the successful treatment of tuberculosis (TB). The study aimed to characterise the impact of pharmacokinetics of 1st-line antituberculosis drugs, and TB and HIV disease markers on TB treatment outcomes (time-to-stable culture conversion (TSCC), relapse, and death).
Methods: TB treatment outcomes were available for 150 patients who participated in the PK sub-study of the RAFA randomised clinical trial (PACTR201105000291300). The RAFA study was a three-arm trial, and patients were randomised to receive (a) efavirenz-based ART at 2 weeks after the start of standard TB treatment, (b) standard TB treatment or (c) 50% higher dose of rifampicin. Steady-state individual AUC0-24 and Cmax were derived from population PK models for rifampicin, isoniazid, pyrazinamide, and ethambutol. Sputum was collected bi-weekly in the first 8 weeks of treatment and monthly thereafter for TB culture. TSCC was defined as the time of the first observed culture negative result which was confirmed at a follow-up visit after 2 or 4 weeks. Long-term treatment outcome was defined as either treatment relapse or death in the 24 months of follow-up. Binary classification and regression trees (CART) and time-to-event (TTE) modelling were used in combination to explore this dataset. CART with conditional inference using permutation tests was applied to identify the most promising covariates effects that are predictive of treatment outcomes and interactions between these covariates (1, 2). The conditional inference methodology eliminates the bias toward selection of continuous predictors associated with ordinary CART. The result of CART analysis is a tree-like structure with binary splits and the root node represents the most significant predictor of the outcome. Additional significant covariates are added as daughter nodes in order of importance using a predefined type 1 error (α = 5%). The identified nonlinear associations in the TSCC analysis were further evaluated using TTE pharmacodynamic modelling with interval censoring. The analyses were implemented in R and Monolix 2016R1 software (3, 4).
Results: TSCC was associated with chest X-ray grading, WHO HIV stage before TB diagnosis, and study arm. TTE-CART identified X-ray grading as the most important covariate associated with TSCC: patients with at most advanced X-ray grade were separated from patients with very advanced X-ray grade (median TSCC = 22 vs 29 days respectively, p value= 0.004). Among patients with very advanced X-ray grade, patients classified as HIV stage 2 before TB diagnosis had faster TSCC compared to those in stage 3 or 4 (median TSCC = 21 vs 29 days respectively, p value= 0.048). In the last step, patients with advanced HIV disease, initiating ART at 2 weeks was associated with faster TSCC compared to standard treatment or high dose rifampicin (median TSCC = 28 vs 36 days respectively, p value= 0.043). Further analysis using TTE pharmacodynamic modelling showed that patient with very advanced X-ray grading, advanced HIV stage, and who did not start ART at 2 weeks converted later than the rest of the cohort (median TSCC: 36 vs 23 days, p value<0.001). Death or relapse was observed in 20/150 patients in the cohort. Patients with lung cavitation were more likely to either relapse or die (25% vs 11%, p value= 0.045). Among patients without lung cavitation, being not physically active (i.e. confined to bed) was associated with relapse or death (24% vs 6, p value= 0.021). The CD4+ count was associated with treatment outcome in patients who did not present with lung cavitation and were physically active. A higher proportion of poor treatment outcome was observed in patients who had lower (< 100 cells/microL) CD4+ count (17% vs. 2%, p-value=0.006)
Conclusions: We combined time-to-event regression trees and pharmacodynamic modelling techniques to describe the relationship between tuberculosis treatment outcome vs drug exposure, disease burden, and patient characteristics. Drug exposure was not found to be associated with either TSCC or long-term TB treatment outcomes, contrary to previous reports (5). Treatment outcomes were correlated with tuberculosis and HIV related disease severity markers. The identified markers could be used to profile patients with high or low risk of treatment failure and to tailor a treatment strategy based on the patient profile.
References:
[1] Fu W, Simonoff JS. 2017. Survival trees for interval-censored survival data. Stat Med 36:4831–4842.
[2] Hothorn T, Hornik K, Zeileis A. 2006. Unbiased recursive partitioning: A conditional inference framework. J Comput Graph Stat 15:651–674.
[3] Monolix version 2016R1. 2016. A software for the analysis of nonlinear mixed effects models. Lixoft SAS, Antony, France.
[4] R Core Team. 2017. R: A Language and Environment for Statistical Computing. Version 3.1.2. Vienna, Austria.
[5] Pasipanodya JG, McIlleron H, Burger A, Wash PA, Smith P, Gumbo T. 2013. Serum drug concentrations predictive of pulmonary tuberculosis outcomes. J Infect Dis 208:1464–73.
Reference: PAGE 27 (2018) Abstr 8703 [www.page-meeting.org/?abstract=8703]
Poster: Drug/Disease Modelling - Infection