Inside the Pune Pilot: What the Evidence Shows About AI-Guided TB Screening

The AI-NETRA Pune Pilot Study looked at whether artificial intelligence can help tuberculosis programmes find cases more efficiently, and whether diagnostic networks can be planned more fairly. The work took place in Pune district, Maharashtra, with the State TB Office, Maharashtra, and implementation partner FIND, and with development support from Google.org. It was reviewed and approved under a formal ethics protocol. This account sets out the design, the findings, and the limits of the evidence. The full methods, statistics, and appendices are in the complete study report.

How the study was designed

The evaluation had three parts, each answering a different question.

The first part looked backwards. AI-NETRA's predictions were compared with three screening campaigns already carried out in Pune district, in October 2023, December 2024, and March 2025. Only the notification data available before each campaign was used. The question was whether the places the model flagged as high risk were the places where cases were later confirmed.

The second part asked whether the tool was usable. Programme leaders and one hundred frontline workers, across sixteen group discussions, reviewed AI-NETRA's outputs. One hundred and one community health workers, known as ASHAs, were surveyed after using AI-NETRA plans in the field.

The third part looked forward. Over two weeks, teams using AI-NETRA screened households in the places the model ranked highest, across four tuberculosis units: rural Bhor and Shirur, and urban Haveli PMC and Haveli PCMC. Results were compared with the standard screening campaign in the same units.

What the backward look showed

In the March 2025 campaign, AI-NETRA's targeting plan would have found about 89 percent of the cases later confirmed, while directing screening to under one-third of the people actually screened. Under standard programme costing, concentrating screening in this way could reduce the cost of a full district-wide cycle from about Rs 7.3 crore to Rs 3.3 crore.

What the field comparison showed

Across the four units, AI-guided teams screened 66,095 people and confirmed 10 cases of pulmonary tuberculosis that were verified by a laboratory molecular test. The standard campaign screened a much larger group of 177,322 people and confirmed 9 cases under the same definition. Put another way, AI-guided screening needed about 6,610 people screened for each confirmed case, compared with about 19,702 under standard practice, a reduction of roughly two-thirds.

What the evidence does not yet prove

When every confirmed case is counted, not only those verified by the laboratory molecular test, the two approaches did not differ in a meaningful way. The AI-guided arm confirmed 15 cases and the standard arm confirmed 32. A large part of this difference is explained by diagnostic access. The standard campaign had a mobile chest X-ray van. The AI-guided sweep did not. The comparison campaign was also run at a different time of year.

The main finding also rests on a small number of confirmed cases. Removing a single confirmed case would weaken the result. It should be treated as a starting point, not a final answer. The report's recommended next step is a trial in which both approaches run at the same time, under matched conditions and the same case definitions.

Diagnostic network planning

A related piece of work, AI-DNO, looked at Pune's testing network of 49 devices across 43 laboratories, handling 53,705 tests a year. Redesigning how samples are referred, without buying new machines, reduced the average distance a sample had to travel by nearly 25 percent, eased pressure on the district's only over-capacity laboratory, and is projected to save about Rs 22 lakh a year.

Conclusion

AI-NETRA is a usable targeting layer that concentrates screening in neighbourhoods where positivity is higher. It reduced the number of people who needed to be screened, and the cost per confirmed case, by roughly two-thirds. The case for using it as an efficiency layer inside existing programmes is well supported. Whether it finds more cases in total remains an open question, and that is the right focus for a later trial. The full methods, statistics, and appendices are in the Pune Pilot Study report.