Cough biomarkers for diagnosis and monitoring of respiratory disease: a systematic review.

Cough is a common and physiologically informative component of respiratory morbidity, but its potential for diagnosing and monitoring disease is not thoroughly investigated. This systematic review synthesised the literature on algorithmic and statistical models analysing cough acoustics for diagnosing or monitoring respiratory conditions. Following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, five databases (PubMed, Embase, Scopus, Web of Science and CENTRAL) were systematically searched for studies published from January 2010 to June 2025. Eligible studies performed quantitative acoustic feature analysis of human coughs using statistical, machine learning or deep learning models and reported diagnostic or prognostic performance. 89 studies from 34 countries were assessed, covering cough detection (n=31), disease classification (n=55) and disease severity prediction (n=3), reflecting potential applications in disease monitoring. Deep learning approaches, especially convolutional and recurrent networks, were predominant (n=56) and tended to achieve the higher accuracies, although machine learning ensemble methods and logistic regression also demonstrated strong performance, particularly with well-engineered features. Across different diseases, sensitivities and specificities were often reported to be ≥90%, notably for tuberculosis, asthma and COVID-19. However, methodological weaknesses were common, with only 11.2% of studies introducing external validation, 71.1-87.6% demonstrating high risk of bias (according to PROBAST-AI) and most based on small, homogeneous or crowdsourced cohorts with limited generalisability. These limitations contribute to inflated internal performance and uncertainty about real-world applicability. Cough acoustic biomarkers hold promise as an adjunctive tool for screening and longitudinal monitoring in low-resource environments. Nevertheless, widespread implementation will require large, multicentre validation, standardised calibration, bias control and incorporation into privacy-preserving workflows.