A smartphone app will analyse the sound of a person’s cough to diagnose tuberculosis, asthma, and COPD in rural and semi-urban Tanzania, where diagnostic equipment is scarce. This matters because Tanzania has one of the world’s highest TB burdens, and non-communicable respiratory diseases are rising. Most primary health facilities lack the tools to diagnose these conditions, and patients cannot afford travel to distant diagnostic centres. The project will recruit 30,000 participants from primary care settings, building an Africa-specific database of cough sounds linked to verified diagnoses. An AI model will then be trained to distinguish TB, asthma, and COPD from cough audio alone, and the resulting classifier will be deployed via a smartphone app and web interface for use by community health workers. If successful, the Kikohozi classifier could enable rapid, non-invasive diagnosis and treatment monitoring at the point of care, reducing disease spread and premature death. The project also builds AI development capacity in Africa, with potential for scaling across sub-Saharan Africa.
View original technical description
The Kikohozi classifier project aims to enhance respiratory disease diagnosis and treatment monitoring in Tanzania through an artificial intelligence (AI)-powered cough audio classifier tailored for resource-limited settings. Named after the Swahili word for "cough," the classifier will analyze cough sounds to identify TB, asthma, and COPD, supporting rapid, non-invasive diagnosis and treatment tracking. Accessible via a smartphone app and web interface, the Kikohozi classifier will empower health workers in community and primary care settings to assess and monitor patients effectively. The web interface will assist health professionals in patient management. There is potential for future use by the public. Designed in, and for, Africa, Kikohozi classifer will address local challenges while advancing AI development capacity in Africa. Our study will be undertaken in Tanzania, which has one of the highest burdens of TB globally and an increasing prevalence of non-communicable respiratory diseases such as asthma and COPD. We will recruit 30,000 study participants from semi-urban and rural areas through the mixed-economy network of primary healthcare facilities. These resource-constrained facilities lack diagnostic capacity for respiratory conditions, and patients lack resources for travel to diagnostic centers. Most health workers own a smartphone. The project supports SDG 3.3 by enabling rapid TB assessment in communities and primary care, facilitating early diagnosis and treatment to reduce the spread and improve outcomes. It advances SDG 3.4 by aiding early detection of asthma and COPD and monitoring treatment response, reducing premature mortality from non-communicable diseases. Aim: To develop and test the Kikohozi classifier and evaluate the feasibility of use in primary care in semi-urban/rural Tanzania. Objectives: Collect and curate an Africa-based database of cough sounds and associated diagnoses (Kikohozi DB) from 30,000 primary care patients, comprising 5,000 each for verified TB, asthma, and COPD, and 15,000 from individuals without these conditions (WP1). Analyze the Kikohozi DB and develop an AI model (Kikohozi-AI) for disease identification, progression, and treatment effects (WP2). Develop a smartphone app (Kikohozi App) and an associated web interface (Kikohozi Web) for use in routine healthcare (WP3). Establish the feasibility and cost-effectiveness of using the Kikohozi App in primary care in Tanzania and explore opportunities and challenges for scale-up (WP4). Methods: WP1: Data Collection for Kikohozi-DB Cough sounds, demographic, and medical data will be collected from individuals with and without respiratory conditions in semi-urban and rural settings through healthcare facilities and community outreach, leveraging partnerships with the Ministry of Health and Aga Khan Health Services. WP2: AI/ML Model Development for Kikohozi App Cough audio data will be pre-processed, features extracted, and recordings augmented for training diversity. Advanced foundation models will be fine-tuned, evaluated with performance metrics, and tested for robustness across devices and demographics to ensure real-world adaptability and iterative refinement. WP3: Applications Development for Kikohozi App & Web A user-centered design approach will guide the development of smartphone and web platforms. Stakeholder input will inform app features and deployment strategies, integrating machine learning (ML) models for real-time diagnostics, secure data management, and telemedicine. Rigorous testing will ensure usability and data compliance. WP4: Health System Pathway Integration Kikohozi App implementation will be piloted in 20 primary care facilities for six months, with data collection, observations, and interviews informing a refined implementation manual and policy recommendations. Working with policy makers, we will plan upscaling for the use of the Kikohozi Classifier App in Tanzania/sub-Saharan Africa.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know