Predictive Modeling and Machine Learning Insights into Multi-Drug Resistant Tuberculosis Treatment Outcomes

Tuberculosis (TB) remains a major global health burden, particularly in resource-prevented settings. Predictive models may improve patient outcomes by identifying individuals at higher risk of treatment failure, but many existing models are trained on data from high-income countries and may not generalize well to low- and middle-income contexts. In this study, we analyzed two datasets: a longitudinal, multi-site cohort, termed the International Collaborations in Infectious Disease Research (ICIDR), with data from Tanzania, Bangladesh, and Siberia and a cross-sectional dataset from Kibong’oto Infectious Disease Hospital (KIDH) in Tanzania. We applied logistic regression, random forest, and XGBoost models to predict outcomes in multidrug-resistant TB (MDR-TB). For the KIDH dataset, patients were stratified by treatment regimen (bedaquiline, short-term injectable, and long-term injectable) to capture regimen-specific predictors. Models trained on the ICIDR dataset showed limited predictive power, likely due to data sparsity and irregular follow-up. In contrast, stratified XGBoost models trained on the KIDH dataset performed better, with F1 scores of 0.766, 0.667, and 0.787 for the respective regimens. BMI consistently emerged as a top predictor, underscoring the role of nutrition in TB recovery. These results suggest that stratifying predictive models by treatment regimen improves both performance and interpretability. However, small sample sizes and inconsistent data collection remain key limitations. Future studies should prioritize standardized, prospective data collection and interdisciplinary collaboration to develop robust, contextually relevant predictive tools.