Noise-Robust TB Cough Screening with HuBERT
A research project on how background noise affects tuberculosis cough classification. Frozen HuBERT embeddings feed a patient-level classifier that I tested across six noise levels and compared with MFCC + SVM and Random Forest baselines.
Evaluation used patient-level cross-validation with bootstrap confidence intervals (121 patients, about 10,000 cough clips), a label-permutation check (mean AUROC 0.49), and an audit of cohort-label confounding. On clean audio the model reached an AUROC of 0.71; noise augmentation lowered that slightly (0.68) but kept far more sensitivity in heavy noise. These are exploratory results, not clinical claims.