A71-03 Early Bedaquiline Adherence Predicts Six-Month Treatment Outcomes in People Living With HIV and Multidrug-Resistant Tuberculosis Using a Time-Aware Machine Learning Model

Rationale

Despite shorter, all-oral regimens, treatment success for rifampicin-resistant and multidrug-resistant tuberculosis (RR/MDR-TB) remains low, especially among people living with HIV. Early identification of patients at risk of poor outcomes is critical for timely intervention. We hypothesized that early adherence data from electronic dose monitoring would enhance the prediction of end-of-treatment outcomes in individuals with RR/MDR-TB and HIV.

Methods

We prospectively studied adults with RR/MDR-TB and HIV initiating bedaquiline-based therapy in KwaZulu-Natal, South Africa (N = 282). Bedaquiline adherence was recorded daily using Wisepill RT2000 electronic medication devices and summarized weekly over 24 weeks. A time-aware extreme gradient boosting (XGBoost) framework incorporated cumulative adherence through each week (1-24), using lagged 3-week rolling summaries (mean, variance, slope) and baseline sociodemographic and clinical covariates. Models were trained with inverse-probability-of-censoring weights to address survivor bias, and hyperparameters were optimized via 10-fold cross-validation. Model discrimination (AUROC, AUPRC) and calibration were evaluated at each weekly horizon to identify the earliest accurate prediction point.

Results

Overall, out of 282 participants, 26% (n = 73) experienced negative RR/MDR-TB outcomes, which included death (n = 43, 15%), loss to follow-up (n = 27, 10%), or treatment failure (n = 3, 1%). Model discrimination improved steadily as the number of adherence observations increased, indicating greater predictive value with each additional observation. At baseline, using sociodemographic and clinical variables as predictors, moderate discrimination was observed (AUROC=0.77; 95% CI, 0.64-0.87). The incorporation of weekly adherence features led to consistent improvements in model accuracy and calibration. By week 4, the model achieved an AUROC of 0.80 (95% CI, 0.69-0.90), sensitivity of 0.80, and specificity of 0.74. Model discrimination continued to improve through week 8 (AUROC=0.89, 95% CI 0.80-0.95), demonstrating stable calibration and a strong positive predictive value, but week 4 represented the earliest actionable horizon, balancing timeliness with reliability. Week-4 adherence was strongly correlated with subsequent adherence intervals (week 5-24, ρ = 0.70, p < 0.001), confirming persistence of early adherence patterns.

Conclusion

Early bedaquiline adherence measured within the first four weeks of treatment provides a robust and clinically meaningful signal of end of treatment outcomes among individuals with RR/MDR-TB and HIV. Integrating electronic dose-monitoring data into a time-aware predictive model enables early, individualized risk stratification at a stage when targeted adherence and clinical interventions can meaningfully alter disease trajectories. Incorporating this approach into routine TB programs has the potential to support proactive, data-driven care, improve resource allocation, and improve treatment success in high-risk populations.

This abstract is funded by: the Stony Wold-Hurbert Fund (PT24-4036), the American Thoracic Society (PG014635-01), and the Burroughs-Wellcome Fund/Ravason Foundation (1467160), National Institutes of Health (2T32HL105323-06A1)