LLM-derived embeddings outperform tabular models for end-of-therapy outcome prediction in tuberculosis patientsNew Data Models Improve Prediction of Tuberculosis Treatment Success

Relapse after apparently successful tuberculosis (TB) therapy remains difficult to predict, and how relapse risk evolves throughout treatment remains unclear. Using harmonised clinical data from two Phase 3 trials (2,918 participants), we performed time-resolved prediction of end-of-therapy (EOT) outcomes and post-treatment relapse by training models using tabular data at monthly intervals from baseline to therapy end. Prediction of EOT outcomes improved after month 3 (ROC-AUC up to 0.84), driven by sputum-smear and solid culture. In contrast, relapse prediction among participants with favourable EOT outcomes and completed follow-up improved only modestly through month 3 (ROC-AUC 0.58-0.63) before declining, with age, sex, clinical symptoms and bacterial burden contributing most strongly to prediction. Models trained on large language model-derived embeddings, a more flexible representation of the same variables, matched tabular relapse models in performance throughout therapy, and outperformed tabular EOT outcome models at months 3 and 4 ({Delta}ROC-AUC: 0.12 and 0.14), with longitudinal modelling and sparse variable inclusion only improving relapse prediction at months 4-6 ({Delta}ROC-AUC: 0.09, 0.19 and 0.12), however with reduced interpretability. Models incorporating data after baseline provided incremental improvements in post-treatment relapse risk stratification compared with baseline cavitation and sputum smear alone (4-month relapse-free survival: 94.8%/78.5% for model-derived low/high risk groups, vs. 91.4%/82.9% for baseline easy-/hard-to-treat groups). Overall, these findings suggest that while routine clinical data collected after baseline can improve post-treatment risk stratification, it offers limited predictive value for relapse, underscoring the need for relapse-specific biomarkers.