AI Utilized for Global Tuberculosis Detection
At the Boniaba Community Health Center in Mali, a recent tuberculosis (TB) screening identified a positive case within seconds. This process occurred without the immediate presence of a physician, representing a shift from previous methods that involved prolonged waiting periods for laboratory results. Previously, TB screenings were less accessible, and sputum test results could take one to two weeks to be processed by a laboratory. The current methodology involves mobile X-ray machines combined with artificial intelligence (AI) algorithms for TB detection.
Global TB Challenge and AI Solution
Tuberculosis is classified as the leading infectious disease cause of death globally, accounting for over 1.2 million deaths annually, equivalent to approximately 3,500 deaths per day. Incidence rates are increasing. A challenge in TB control has been a worldwide scarcity of radiologists needed for diagnosing this bacterial infection, which primarily impacts the lungs. Dr. Lucica Ditiu, executive director of the Stop TB Partnership, stated that some countries have fewer than five radiologists, typically concentrated in capital cities. More than 80 low- and middle-income countries are now implementing AI for TB screening. Ditiu characterized this development as revolutionary.
She cited an example where AI-enabled X-ray screenings are being used among nomadic populations in remote areas of Nigeria. The Stop TB Partnership contributed to the early development of this technology eight years prior. AI models are additionally deployed in refugee camps in Chad, where radiologists are absent. Peter Sands, Executive Director of the Global Fund to Fight AIDS, TB and Malaria, noted the AI's role in interpreting X-rays. The Global Fund has invested nearly $200 million in AI-enabled TB screenings over the last four years.
Operational Details and Impact
Advocates of the technology suggest AI will enhance global disease detection and control, particularly in underserved communities. Conversely, some stakeholders recommend additional regulations and safeguards to protect patients in low- and middle-income countries. At the Boniaba Community Health Center, Diakité Lancine, a trained X-ray technician, operates a mobile X-ray machine. Images captured are sent to a computer, where an AI model analyzes them and generates a score indicating the likelihood of TB, along with a visual representation of lung anomalies. Lancine explained that blue areas on the visual representation indicate normal findings, while red areas suggest abnormalities.
Lancine, employed by the nonprofit ARCAD Santé PLUS, conducts TB screenings across West Africa, utilizing portable equipment including a mobile X-ray machine, computer, and battery pack. Following a screening with red indicators, Lancine collected a sputum sample for laboratory confirmation and advised the individual to bring their children for screening, due to TB's airborne transmission risk within households. AI analysis identified potential TB in three of the children. Subsequent antibiotic treatment for six months was planned.
Bassy Keita, program officer at ARCAD Santé PLUS, stated that AI has significantly impacted screenings. He noted the difficulty of obtaining sputum samples from children. The introduction of AI screenings has enabled the rapid identification of individuals without X-ray indications of TB, reducing the number of required sputum samples by approximately 50% for those with potential TB.
TB at the Forefront of AI Solutions
Regina Barzilay, a professor and computer scientist at MIT, developed AI models for breast and lung cancer detection. She subsequently created a TB screening AI model for a Sri Lankan hospital that could not afford commercial alternatives. Barzilay identified TB as a key area for AI-driven global health solutions. She explained that TB's visual manifestation on X-rays, combined with existing diagnostic labels, facilitates straightforward and cost-effective AI model training, estimating development time at a few months and under $50,000. TB X-ray machines are widely available in low-resource settings and require minimal operator training, contrasting with equipment for mammograms or blood tests.
Ditiu highlighted the significant need, noting that the World Health Organization reported 10.8 million new TB cases in 2023, an increase from 10.1 million in 2020, with most cases occurring in low- and middle-income countries. Ditiu suggested that TB is an initial application, as some existing AI models developed for TB can also diagnose conditions such as lung cancer, pneumonia, and certain cardiovascular issues. Barzilay predicted that AI integration into health systems in low-income countries could parallel the rapid adoption of mobile phones over landlines in Africa. She stated that AI adoption might be faster in developing countries due to unmet needs and clinicians' recognition of assistance requirements, noting that while much of the technology is developed in the United States, its application is global. Barzilay attributed slower AI integration in healthcare in countries like the U.S. to challenges in widespread utilization and incorporation into professional care guidelines, even for FDA-approved models.
Challenges and Concerns
However, some experts advise caution regarding rapid AI adoption in healthcare. Erwin John Carpio, a radiologist in the Philippines, assisted in developing AI guidelines for the Philippines College of Radiologists and researched AI use for TB screening in remote provinces. Carpio noted that many high-income countries possess regulations for AI in health, whereas developing nations, often offered the technology for free, face challenges in establishing similar safeguards. He raised concerns about AI models failing to diagnose TB, leading to missed medical attention.
Carpio stated that the U.K. and U.S. have systems for reporting AI diagnostic errors and improving patient safety, often linked to regulatory approval processes like the FDA, but such laws are not yet established in the Philippines. He also expressed concern about AI model drift, where performance degrades silently over time without user notification of decreased accuracy. Mitigation strategies for AI errors include training models to flag complex cases and implementing continuous quality control checks by external experts, as observed in Lancine's Mali program and Global Fund-supported initiatives.
Carpio noted that comprehensive quality control requires a team of specialists, including radiologists, computer scientists, data scientists, and AI engineers. He suggested that the combined cost and energy consumption for such operations might exceed initial perceptions of affordability and simplicity. Advocates argue that AI's efficacy should be compared against existing alternatives. Barzilay stated that medical errors committed by physicians are common. Sands, from the Global Fund, indicated that in many deployment environments, the scarcity of radiologists for TB diagnoses means AI offers a viable solution in their absence. Sands referenced data suggesting improved global TB detection rates since the World Health Organization endorsed AI technology in 2021 and provided a toolkit for local calibration. Barzilay concluded with the question of whether adequate medical care would be available for all individuals diagnosed via AI.