AI-NETRA Pune Pilot Study: Field Evidence for TB Case Finding
A three-part mixed-methods pilot evaluation of AI-guided active case finding and diagnostic network optimization in Pune district, Maharashtra.
Document Information
- Government partner
- State TB Office, Maharashtra
- Implementation partner
- FIND
- Development partner
- Froncort.AI with support from Google.org
- Ethics approval
- JCDC/BHR/26/005, 13 March 2026
Abstract
Background. Tuberculosis active case finding under India's National TB Elimination Program is effective but inefficient when screening is untargeted. TB clusters spatially, creating an opportunity for geographic targeting. AI-NETRA predicts neighborhood-level TB risk on an H3 hexagonal grid from Nikshay notification data and translates predictions into field-ready microplans.
Methods. A three-part mixed-methods evaluation in Pune district: retrospective validation against three ACF campaigns; acceptability and usability with program leadership and 100 frontline workers; prospective field comparison of AI-guided versus standard ACF across four TB Units.
Results. Retrospectively, AI-NETRA captured 89% of notified cases from under one-third of the screened population and reduced NNS by 64%. Prospectively, the AI arm confirmed 10 NAAT-positive pulmonary cases from 66,095 screened versus 9 from 177,322 in the standard arm. NNS was 6,610 versus 19,702; cost per bacteriological case fell from Rs 27.6 lakh to Rs 9.3 lakh.
Conclusions. AI-NETRA is a usable targeting layer that concentrates ACF into higher-positivity pockets and lowers NNS and cost per confirmed case. A concurrent, like-for-like trial with harmonized case definitions is the appropriate next step.
Headline Results
Retrospective validation: cases found from under one-third of the screened population (March 2025 campaign)
Prospective rate ratio vs. standard ACF
Number needed to screen: 6,610 vs. 19,702 under standard campaign
Down from Rs 27.6 lakh per bacteriologically confirmed case (66% reduction)
AI-DNO referral redesign reduced average service distance with existing machines
4.3% of tested presumptives confirmed under AI targeting vs. 1.2% under standard ACF
What the Evidence Does and Does Not Yet Show
Total case detection did not differ
Using an all-forms definition, the AI arm confirmed 15 cases and the standard arm 32. The standard campaign had access to a mobile X-ray van; the AI sweep did not, so case ascertainment differed between arms.
Primary endpoint rests on small counts
The prospective bacteriological result is promising as observed but rests on a small number of confirmed cases, so it should be treated as a starting point rather than a final answer.
Non-concurrent comparator
The standard arm was a campaign from a different time period (December 2024–January 2025). Secular trends and seasonal effects are not controlled.
AI-DNO concordance below threshold
Model recommendations reached 77.1% concordance with an expert panel, below the pre-specified 80% acceptance threshold. Efficiency gains are real; concordance is promising but not yet meeting the bar set for it.
Methods at a Glance
Part 1: Retrospective validation
Three successive ACF campaigns in Pune district (October 2023, December 2024, March 2025) compared against temporally matched AI-NETRA predictions using only pre-campaign Nikshay notification data.
Part 2: Acceptability and usability
Program leadership survey (n=15), 16 focus group discussions with 100 frontline workers, controlled time-and-motion study (n=57), and post-campaign ASHA survey (n=101) after real field use.
Part 3: Prospective field evaluation
Two-week AI-guided ACF sweep across four TB Units (Bhor, Shirur, Haveli PMC, Haveli PCMC) compared against the corresponding standard campaign. Primary endpoint: bacteriologically confirmed, NAAT-positive pulmonary TB.
Sub-studies
AI-DNO diagnostic network optimization across 49 devices, 43 laboratories, and 53,705 tests; PDFM cold-start prediction evaluation in zones without notification history.
Reporting follows STROBE, TREND, TRIPOD+AI, and COREQ standards.
Acceptability and the Field
Field photographs used with consent. AI-NETRA prospective deployment, Pune district, 2025–2026 (Bhor, Shirur, Haveli PMC, Haveli PCMC).
Limitations
- Non-concurrent comparator: standard arm from a different time period.
- Different case ascertainment: mobile X-ray van available in standard arm but not AI arm.
- Different urban denominators: population-covered vs. household-enumeration bases.
- Implementation bundle: results reflect AI-NETRA plus training plus microplan workflow, not the algorithm alone.
- Fragile primary endpoint at small case counts (10 NAAT-positive pulmonary cases in AI arm).
- Equity and cold-start dependence: predictive advantage weakest where notification history is sparse.
- AI-DNO concordance (77.1%) below pre-specified 80% threshold.
- PDFM cold-start sub-study limited to a single city and year; directional only.
Way Forward
- Concurrent, like-for-like cluster-randomized trial with harmonized case definitions and equal mobile X-ray deployment.
- Environmental cold-start layer using weather and air-quality signals for zones without notification history.
- Offline, field-first local-language application with voice navigation and automatic geocoded capture.
- Program enablers: timely ASHA incentives, data-cost reimbursement, and reliable mobile X-ray access.
Read the full study report
Complete methods, results, and appendices — 22 July 2026