Geographical targeting of active case finding for tuberculosis in Pakistan using artificial intelligence software: a qualitative study embedded within the SPOT TB trial

Tuberculosis remains one of the world's most persistent killers, a disease that thrives in the shadows of crowded cities and remote villages alike. In countries like Pakistan, where the burden of the disease is exceptionally high, health workers face a daunting challenge: finding the millions of people who are sick but have not yet been diagnosed. To catch these hidden cases, programs send mobile teams into communities to set up screening camps, using X-ray machines to look for signs of infection in people's lungs. For years, the decision of where to set up these camps relied on human judgment, local maps, and the experience of field staff. But as the volume of data grew, health officials began to wonder if computers could do a better job. They turned to artificial intelligence, a technology capable of sifting through vast amounts of information to spot patterns humans might miss. The hope was that a smart computer program could point health workers directly to the neighborhoods where sick people were most likely to be found, making the search faster and more effective.

This question drove a large-scale experiment in Pakistan known as the SPOT-TB trial. Researchers wanted to see if using an artificial intelligence tool called MATCH-AI to choose camp locations would find more tuberculosis cases than the traditional method of picking sites based on local knowledge. The tool analyzed routine health data and local conditions to generate a list of recommended locations. However, a computer's recommendation is only as good as the reality on the ground. To understand why the trial did not show a clear overall improvement in finding cases, a separate team of researchers conducted a deep dive into the human side of the project. They spoke with the doctors, supervisors, and coordinators who actually drove the vans and set up the tents, asking them about their daily experiences with the new technology. Their goal was not just to see if the software worked on a screen, but to understand how it functioned in the complex, messy reality of Pakistan's health system.

The researchers found that the artificial intelligence tool was a double-edged sword. On one hand, many field staff appreciated the software for bringing a sense of fairness and structure to their work. Before the tool, the choice of where to hold a camp could be swayed by personal connections, political pressure from local leaders, or the simple habit of returning to the same familiar spots. The software removed this human bias, generating a list of sites based purely on data. This helped staff resist pressure from officials who wanted camps in specific areas for their own reasons. In some instances, the tool pointed teams toward remote or previously overlooked villages where they discovered clusters of tuberculosis cases that would have otherwise gone undetected. One team, for example, followed a recommendation to a distant household in Faisalabad and found five infected people living together, a discovery they credited to the software's ability to look beyond the usual boundaries.

Yet, the story of the tool's success was complicated by the harsh realities of the landscape and the limits of the technology itself. The software, while good at crunching numbers, lacked the common sense that comes from living in the community. It sometimes suggested locations that were physically impossible to reach, such as snow-covered mountain peaks or areas cut off by rivers, without realizing that a mobile van could not drive there. In other cases, it pointed to places that were unsafe due to local security risks or areas where cultural norms made it impossible for women to travel far or visit a public camp. In conservative regions, for instance, women might refuse to leave their homes to travel five kilometers to a camp in a busy market, especially if the screening area lacked privacy. The software did not understand these social barriers; it only saw a dot on a map. Field staff, who knew the terrain and the people, often had to override the computer's suggestions, using their own judgment to find spots that were accessible and acceptable to the community.

The gap between the digital plan and the physical world was widened by the infrastructure needed to run the system. The software relied on a digital platform to collect data in real time, but in many remote areas, the internet connection was unreliable or non-existent. When the signal dropped, the system would freeze, or data entered in the field would fail to sync with the central server. This forced health workers to revert to old-fashioned paper records, which they later had to type into the computer when they returned to the city. This double work created frustration and fatigue, slowing down the very process the technology was meant to speed up. Furthermore, the people on the front lines often did not fully understand how the software made its decisions. They were told to follow the list of locations but were rarely given a clear explanation of the logic behind them. This lack of understanding made it difficult for them to trust the tool or feel a sense of ownership over the new system, leading some to ignore its recommendations in favor of their own experience.

The study concluded that while artificial intelligence holds genuine promise for organizing and improving tuberculosis screening, it cannot replace the deep, contextual knowledge of the people working in the field. The technology is a powerful assistant that can reduce bias and highlight areas that might otherwise be ignored, but it is not a substitute for human judgment. It cannot account for the security of a neighborhood, the cultural sensitivities of a community, or the condition of a dirt road. For the tool to work effectively, it must be supported by reliable infrastructure, such as stable internet connections, and it must be introduced with thorough training that helps staff understand not just what to do, but why. The researchers found that the most successful approach was a hybrid one, where the computer provides a starting point, and the human team uses their local expertise to adapt those recommendations to the real world. In the end, the technology is only as strong as the system that supports it and the people who use it.