📝 SummarXiv

Abstract

Background Artificial intelligence (AI) can detect disease-related acoustic patterns in cough sounds, offering a scalable approach to tuberculosis (TB) screening in high-burden, low-resource settings. Previous studies have been limited by small datasets, under-representation of symptomatic non-TB patients, reliance on simple models, and recordings collected under idealised conditions. Methods We enrolled 512 participants at two hospitals in Zambia, grouped as bacteriologically confirmed TB (TB+), symptomatic patients with other respiratory diseases (OR), and healthy controls (HC). Usable cough recordings plus demographic and clinical data were obtained from 500 participants. Deep learning classifiers based on speech foundation models were trained on cough recordings. The best-performing model, trained on 3-second segments, was further evaluated with demographic and clinical features. Findings The best audio-only classifier achieved an AUROC of 85.2% for distinguishing TB+ from all others (TB+/Rest) and 80.1% for TB+ versus OR. Adding demographic and clinical features improved performance to 92.1% (TB+/Rest) and 84.2% (TB+/OR). At a threshold of 0.38, the multimodal model reached 90.3% sensitivity and 73.1% specificity for TB+/Rest, and 80.6% and 73.1% for TB+/OR. Interpretation Cough analysis using speech foundation models, especially when combined with demographic and clinical data, showed strong potential as a TB triage tool, meeting WHO target product profile benchmarks. The model was robust to confounding factors including background noise, recording time, and device variability, indicating detection of genuine disease-related acoustic patterns. Further validation across diverse regions and case definitions, including subclinical TB, is required before clinical use.