62
Research Trials
22
Peer-reviewed publications
17
Clinical Conditions
First study to assess the antitussive effect of a drug using continuous cough monitoring, tracking patients on a Hyfe-enabled smartwatch for five weeks around the start of azithromycin.
External validation of the CODA TB DREAM Challenge cough-only models in an independent Peruvian cohort recorded with the Hyfe Research app, showing reduced performance outside the original training populations.
Pilot cohort study of behavioural cough suppression therapy delivered to small telehealth groups with rolling enrolment, with exploratory objective cough monitoring using Hyfe’s CoughPro app.
26 May 2026
The CODA TB DREAM Challenge produced eleven cough-only AI models for tuberculosis screening, all developed and validated on data from seven countries. This study tested how those models perform on a population none of them had seen: adults presenting with cough at a public hospital and primary health centres in Lima, Peru.
Solicited coughs from 303 participants were recorded with the Hyfe Research app, the same platform used to build the original CODA dataset. That kept recording format and audio length comparable across the two datasets. About a third of participants had microbiologically confirmed pulmonary TB. The eleven models were run unchanged, without retraining or tuning, which mirrors how pre-trained tools are typically deployed in practice.
Performance declined for every model, with AUCs falling from 0.65–0.74 in the original challenge to 0.48–0.62 in Peru. The challenge’s top-ranked model remained the best performer in the new setting. Discrimination was better in participants aged 35 or older, a group closer in age to the original training population. The authors attribute the drop to differences in population characteristics, local respiratory disease epidemiology and recording hardware.
The findings show that cough-based TB models validated in one set of countries cannot be assumed to transfer to others. They highlight the need for external validation, more geographically diverse training data, or locally tailored models. Using a single, standardized recording platform across sites is what made a direct comparison between the two datasets possible.
