Using AI to Spot Hidden Tuberculosis Risks in Kampala Households

Tuberculosis (TB) remains a persistent health challenge in Uganda, especially in Kampala where the disease overlaps with a high rate of HIV/TB coinfection. Hospital patients in the city almost always show clear TB symptoms, yet active community screening finds that 30% or more of undiagnosed TB cases are asymptomatic — people who carry the infection but show no obvious signs. That contrast, together with a situation where host risk factors cannot be easily separated from environmental ones, makes tracing and stopping TB transmission difficult. To tackle this, a research team led by David Patrick Kateete is applying health data science to an old but still unresolved problem. The project, approved by the School of Biomedical Sciences IRB (protocol number SBS-2023-495), will draw on demographic, clinical and laboratory records from TB patients and their household contacts and use the computing resources at Makerere University. Rather than a single test or clinic study, the work uses data-driven methods to look for patterns that conventional approaches struggle to explain, with the goal of improving early detection and understanding how TB spreads and persists in households.

The core methods of the study are Artificial Intelligence (AI) and Machine Learning (ML) algorithms built from existing health data. The team will combine demographic, clinical and laboratory data and run them on Makerere University computational resources to train models that recognize who is most likely to need extra attention. Specifically, the algorithms will try to identify, at baseline (month 0), patients who are unlikely to convert their sputum or culture results by months 2 and 5 and therefore may be at risk of failing TB treatment. A second aim is to single out household contacts who are at risk of developing TB disease and to identify contacts who may resist TB infection despite repeated exposure to M. tuberculosis. The study protocol describes developing these predictive tools rather than reporting finished results; the work is intended to show whether AI/ML can uncover signals in routine TB data that predict transmission and treatment outcomes.

If the planned AI and ML models can reliably flag patients unlikely to clear infection by month 2 or month 5, and can identify household contacts at higher or lower risk, the practical payoff could be substantial. Early identification of people at risk of treatment failure would allow clinicians and public health teams to prioritize adherence support, closer monitoring, or targeted investigations. Recognizing contacts who resist infection despite repeated exposure could point to protective factors that deserve further study. Applied in a city with high HIV/TB coinfection, the approach could help focus scarce resources on the households and individuals most likely to drive ongoing transmission. The researchers frame this work as a test of whether modern data science methods can provide actionable insights for an ancient disease, offering evidence that such tools can help reduce TB transmission in the community when paired with local health systems and laboratory data.