Time-resolved predictability of end-of-therapy outcome and relapse after cure in Phase 3 tuberculosis trials

The study performed time-resolved prediction of EOT outcomes and post-treatment relapse, showing EOT outcome prediction improved after month 3 (ROC-AUC up to 0.84), driven by sputum-smear and solid culture. How relapse risk evolves throughout treatment was previously unclear; monthly modelling characterises how predictive ability changes over the course of therapy. Based on harmonised clinical data from two Phase 3 trials (2,918 participants), with monthly tabular models evaluated by ROC-AUC.

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. In contrast to EOT outcome prediction, this indicates limited predictive value of routine clinical data for relapse. Time-resolved modelling in the same cohort, with explicit ROC-AUC ranges and ranked variable contributions.

Models trained on large language model-derived embeddings matched tabular relapse models and outperformed tabular EOT models at months 3 and 4 (ΔROC-AUC 0.12 and 0.14), while longitudinal modelling and sparse variable inclusion improved relapse prediction only at months 4-6 (ΔROC-AUC 0.09, 0.19 and 0.12) with reduced interpretability. Introduces a more flexible representation of the same variables and quantifies gains from longitudinal modelling and sparse variable inclusion by month. Gains reported as ΔROC-AUC, with reduced interpretability noted.

Models incorporating post-baseline data provided incremental improvement in post-treatment relapse risk stratification versus 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). Provides a direct comparison between model-derived risk groups and baseline groups. Presented as 4-month relapse-free survival by risk stratum.

The results apply to participants in two Phase 3 trials who received therapy and completed follow-up, using routine clinical variables from baseline to therapy end; for participants with favourable EOT outcomes, model-derived risk stratification can inform discussion of post-treatment monitoring intensity, while the limited relapse prediction points to the need for relapse-specific biomarkers to support finer individualised decisions.

Relapse prediction is limited and declines over therapy, and the clinical utility of model-derived risk groups still requires prospective validation; LLM-embedding models have reduced interpretability, and gains from longitudinal modelling and sparse variable inclusion are concentrated in specific months; additionally, this is an abstract-level reading, so details in figures and supplementary materials are not included, which may affect a complete understanding of variable contributions and model specifications.