Accelerating Cough-Based Algorithms for Pulmonary Tuberculosis Screening: Results From the CODA TB DREAM Challenge

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Public Health

Background:

• Open-access data challenges accelerate innovation in AI-driven health tools.
• The Cough Diagnostic Algorithm for Tuberculosis (CODA TB) DREAM Challenge focused on developing AI for tuberculosis screening.
• Cough sound analysis offers a novel approach for detecting tuberculosis.

Purpose of the Study:

• To develop and validate artificial intelligence algorithms for tuberculosis screening using cough sounds.
• To compare the performance of models using cough sounds alone versus cough sounds combined with clinical data.
• To assess the feasibility of a rapid, open-access data challenge for advancing diagnostic tools.

Main Methods:

• Data from 2143 adults with persistent cough across seven countries were collected.
• Participants provided cough recordings via smartphone and underwent standard tuberculosis evaluation.
• Machine learning models were developed using training data and validated on independent test sets, evaluated by AUROC and pAUROC.

Main Results:

• Eleven cough-only and six cough-plus-clinical models were submitted.
• Cough-only models achieved a maximum AUROC of 0.74 with 55.5% specificity at 80% sensitivity.
• Incorporating clinical data improved AUROCs (0.78-0.83), with 5/6 models meeting the target pAUROC and achieving 73.8% specificity at 80% sensitivity.

Conclusions:

• The CODA TB DREAM Challenge rapidly facilitated the development of AI-based tuberculosis screening algorithms.
• Combining cough sounds with clinical data significantly enhanced diagnostic performance.
• These AI tools show considerable potential as a non-invasive tuberculosis screening method.