← News

Models & toolsSources: arXiv · 18 Nov 2026

Cough-based TB screening models fail when tested on outside datasets, study finds

A systematic evaluation reports that machine learning classifiers for tuberculosis built on cough sounds perform well within their own datasets but drop sharply on external data, with recording device and dataset shaping the audio features.

In plain words

Researchers tested whether software that detects tuberculosis from the sound of a cough works on data it was not trained on. It mostly did not: performance fell when models were applied to other datasets. The audio patterns the models learned reflected the recording equipment and dataset more than the disease itself.

What was announced

The authors evaluated classical machine learning and deep learning classifiers for tuberculosis screening using cough audio across three independent datasets. Within a single dataset, performance was moderate, with ROC-AUC reaching 0.76 plus or minus 0.06, but external performance was frequently below 0.6, which the authors attribute to a possible limitation of the data rather than the modeling approach alone.

The analysis suggests why: audio representations clustered by recording device and dataset rather than by tuberculosis status, and predicted tuberculosis probability tracked country-level prevalence in the CODA dataset. Device mismatch degraded transfer between datasets, while training on device-diverse data improved it. A baseline using clinical variables instead of audio generalized more consistently, with ROC-AUC between 0.65 and 0.71, leading the authors to conclude that acquisition-specific variability in cough acoustics is a stronger driver of poor generalization. They stress that strong within-dataset results are not sufficient and that external validation is essential before such models are clinically ready.

What it does not show

  • The evidence is the paper's abstract only; full methods, dataset details and complete results are not available for verification.
  • The reported figures come from the authors' own evaluation and have not been independently confirmed.
  • The abstract does not specify the datasets' sizes, patient populations or how the models would perform in a clinical setting.

For security and IT teams

Scope
The finding applies to cough-audio machine learning classifiers for tuberculosis screening evaluated across three datasets in this study, not to deployed clinical tools.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.