Finding Hidden TB for Less: AI X-Ray Reading Alone Beat Adding a Blood Test

Of the 20,023 adults screened for tuberculosis in two districts of Lesotho and South Africa, just 1,545 reported even one of the four classic symptoms. That number is the whole problem with TB in one line. According to the Swiss Tropical and Public Health Institute, nearly half of the people this trial confirmed as having TB did not report a cough, fever, night sweats or weight loss. They felt well enough to go about their lives, while the disease went unfound.

A team led by the Swiss Tropical and Public Health Institute, with SolidarMed in Lesotho and the Human Sciences Research Council in South Africa, tested two ways to find those hidden cases. Both used the same AI; the question was whether adding a blood test helped. The simpler one won: AI software that reads chest X-rays, on its own, found more TB and cost less than the same software paired with a blood test. The trial was published July 1 in The Lancet Infectious Diseases.

What exactly was compared?

The work took place in the Butha-Buthe District of Lesotho and the uMgungundlovu District of South Africa, where household members aged 18 and older could take part. Every participant had a digital chest X-ray read by computer-aided detection software called CAD4TBv7. Then the two strategies were scored in the same person, a design the team calls paired screen-positive:

• AI alone: the software’s reading of the X-ray decided who went on to a confirmatory molecular test (Xpert MTB/RIF Ultra).
• AI plus blood test: people whose X-ray score fell in a middle range also got a point-of-care C-reactive protein test, meant to cut down on expensive confirmatory tests without missing cases.

What did the AI do?

It read each chest X-ray and produced a score for how likely TB was, and that score decided who went on to confirmatory testing. According to Swiss TPH, the World Health Organization recommends this kind of computer-aided detection for TB screening, so the trial asked how best to build a community program around it. It did not compare the software with human readers.

And the answer?

Seventy-three people were found to have TB by at least one approach and had complete data for both. The AI alone identified 69 of them (94.5%). The AI plus blood test identified 60 (82.2%). The cost per case found was US$5,454 with the AI alone and $7,486 with the added test, 37.3% more expensive. A companion cost analysis in PLOS Global Public Health found that the extra cost of the blood test was not offset by savings on confirmatory tests.

Who stands to gain?

Health programs in places with heavy TB burdens and thin budgets, first of all. In this trial, 4,534 participants (22.6%) were living with HIV. A plan with one fewer step could also be simpler for field teams to run. The trial was funded by the European and Developing Countries Clinical Trials Partnership.

“Countries such as Lesotho, which have a high TB burden but relatively limited resources for controlling the disease, can draw important insights from our findings regarding community-based TB screening approaches.” Irene Ayakaka, SolidarMed Lesotho, senior author, via Swiss TPH

Lead author Klaus Reither of Swiss TPH said that CAD-supported chest X-ray screening alone “can be a powerful tool to find TB in communities.” The person who matters most here is the one who never would have walked into a clinic: no cough, no fever, just a shadow on an X-ray that a piece of software noticed in time.