Predicting Drug-Resistant Tuberculosis Treatment Outcomes: A Leakage-Controlled Evaluation of Topological Data Analysis Augmentation

We asked what topological structure adds to a clinical prediction model and how that addition should be measured. Drug-resistant tuberculosis (DR-TB) treatment outcome prediction is the testbed. The contribution is an evaluation design and a bounded null result. Cross-sectional records for 9210 patients from the NIAID TB Portals programme were collated to one row per patient by worst-case aggregation, giving a 99-feature clinical schema. We derived two topological representations. A Mapper graph, built on a UMAP lens over a Gower dissimilarity, supplied four neighbourhood-position features. Local Vietoris--Rips persistent homology on each patient's 26-point neighbourhood supplied degree-0 and degree-1 summaries, vectorised by the best of thirteen catalogued methods. We compared five feature configurations across four classifiers in a pipeline in which every topological step is fitted inside the training partition. A sealed test lockbox of 1776 patients, 463 of them with adverse outcomes, was opened once. We compared discrimination by the area under the receiver operating characteristic curve (AUROC) using paired DeLong tests with Bonferroni correction (corrected threshold 0.0125). No topologically augmented configuration meaningfully improved lockbox AUROC over the clinical-only baseline under any classifier. The best AUROC was 0.7548, achieved by the baseline with XGBoost. Across all sixteen comparisons the upper 95% confidence limit never exceeded +0.014 AUROC. Five comparisons met the corrected threshold. Four favoured the baseline, and one, under the support vector machine, favoured the augmented model by a negligible +0.0005 AUROC.The same features supported neighbourhood-based risk stratification and revealed variation in adverse-outcome composition across patient neighbourhoods: adverse-outcome purity ranged from 5.9% to 30.5% across the fifteen largest Mapper nodes, stably under resampling. The findings support a complementary framing of topological data analysis in clinical prediction. Topological features are best understood not as replacements for clinical predictors but as an analytical lens that reveals geometric structure in the patient population not captured by the clinical feature representation alone.