Tuberculosis Screening from Cough Audio: Baseline Models, Clinical Variables, and Uncertainty Quantification.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Respiratory Medicine

Background:

• Tuberculosis (TB) screening using cough audio is gaining interest.
• Current research lacks standardization in datasets, metrics, and protocols, hindering progress.
• Difficulty in comparing studies impedes evaluation of true modeling advances.

Purpose of the Study:

• To propose a standardized, reproducible framework for automatic TB detection.
• To establish a well-documented baseline for TB prediction using cough audio and clinical data.
• To enable fair comparison of different TB detection methods and reduce methodological variance.

Main Methods:

• Developed an end-to-end reproducible pipeline for TB detection.
• Utilized cough audio recordings and clinical metadata from a multi-country dataset.
• Implemented multimodal fusion, cougher-independent evaluation, and uncertainty quantification.

Main Results:

• Quantified performance for both cough audio-only and fused (audio + clinical data) models.
• Established a strong baseline for TB prediction.
• Reported a consistent suite of clinically relevant metrics for fair comparison.

Conclusions:

• The proposed framework provides a common reference point for TB detection research.
• Standardization is crucial for measuring progress and advancing TB screening technologies.
• This baseline aims to accelerate the development and validation of effective TB diagnostic tools.