Geographical targeting of active case finding for tuberculosis in Pakistan using artificial intelligence software (SPOT-TB): a pragmatic stepped wedge cluster randomized control trial.

This pragmatic stepped-wedge cluster randomized trial in Pakistan demonstrates that an AI-guided targeting strategy (MATCH-AI) significantly improves the yield of bacteriologically confirmed tuberculosis during active case-finding camps compared to conventional site-selection methods, particularly in rural districts and areas with moderate baseline yields.

The Gist

The Big Problem: Finding Needles in a Haystack

Imagine you are trying to find a few specific needles hidden inside a massive haystack. In this story, the needles are people with Tuberculosis (TB), and the haystack is the general population in Pakistan.

For years, health workers have been using mobile vans to go around towns and villages, offering free chest X-rays to anyone who shows up. This is called "Active Case Finding." However, the old way of choosing where to park the vans was a bit like guessing. Health workers would rely on their gut feelings, past experiences, or ask local leaders, "Where do you think we should go?"

The problem is that this "guessing game" often led to vans parking in areas where there were very few needles (few TB cases), wasting time and resources.

The New Idea: A GPS for Disease

The researchers wanted to test a new tool: an AI software called MATCH-AI. Think of this software as a high-tech GPS that doesn't tell you the fastest route to the grocery store, but rather the route to the "hotspots" where TB is most likely hiding.

The software looks at a map of Pakistan and uses data (like population density, income levels, and past TB reports) to predict exactly which neighborhoods are most likely to have undiagnosed TB cases. It then gives the health workers a list of specific GPS coordinates to visit.

The Experiment: A Race Between Guessing and GPS

To see if the GPS was better than the gut feeling, the researchers ran a massive experiment across Pakistan involving 30 mobile X-ray vans and 68 districts.

They used a clever setup called a "stepped-wedge" trial. Imagine a relay race where the runners switch lanes at different times:

1. Phase 1: All 30 vans started by using the old method (guessing/local knowledge).
2. Phase 2: Every month, three vans switched to the new method (using the AI GPS).
3. Phase 3: By the end, all 30 vans were using the AI GPS.

This allowed them to compare the same vans using the old method against the same vans using the new method, while also comparing them to each other.

The Results: It Depends on How You Drive

The study found some interesting results, but with a major catch: The tool only works if you actually follow the directions.

1. The "Intention-to-Treat" Result (The Mixed Bag)
When the researchers looked at all the camps the AI vans ran, the results were just a tiny bit better than the old way, but not enough to be statistically significant.

• Why? The researchers found that the vans didn't always go exactly where the AI told them to. Sometimes, due to traffic, road closures, or local commitments, they parked a few miles away from the recommended spot.
• The Analogy: It's like having a GPS that says "Turn left at the red house," but the driver turns left at the next house because the red house was blocked. You end up in the wrong neighborhood, and you don't find the needles.

2. The "Validated" Result (The Success Story)
The researchers then looked only at the camps where the vans actually parked within 5 kilometers of the AI's recommended spot.

• The Result: In these "perfect adherence" camps, the AI-guided teams found 32% more TB cases than the teams using the old guessing method.
• The Takeaway: When the health workers followed the AI's map exactly, they were much more efficient at finding the disease.

3. Where Did It Work Best?
The AI software was particularly good at finding needles in:

• Rural areas: Places with lower population density.
• "Medium-Yield" districts: Areas where TB exists but isn't obvious. In areas where TB is already everywhere (high-yield), the old methods were already doing a decent job. In areas where TB is very rare (low-yield), the AI couldn't find many needles because there simply weren't many to find.

The Bottom Line

This study proves that AI can act as a powerful compass for finding Tuberculosis. However, the compass is only useful if the drivers actually follow the route.

When health workers in Pakistan used the AI software to pick their locations exactly as instructed, they found significantly more TB cases than when they relied on their own experience. This suggests that in the future, mixing human experience with data-driven AI maps could help save lives by ensuring screening vans go exactly where they are needed most.

Technical Summary

Technical Summary: Geographical Targeting of Active Case Finding for Tuberculosis in Pakistan using AI (SPOT-TB)

Problem Statement
Tuberculosis (TB) remains a critical public health challenge in Pakistan, which accounts for 6.3% of the global TB burden and faces a 25% gap between estimated incident cases and notified cases. While community-wide Active Case Finding (ACF) via mobile chest X-ray "camps" has been scaled up to close this gap, the efficiency of these interventions is diminishing. As screening expands to cover the general population, detection yields decrease. Previous reviews and programmatic data suggest that conventional site selection strategies—relying on field staff experience, historical paper registries, and local stakeholder consultation—are often suboptimal, leading to screening in areas with low TB prevalence and diluting the impact of resources. There is a lack of prospective, randomized evidence evaluating whether data-driven, artificial intelligence (AI) guided geographical targeting can improve TB detection yields compared to routine programmatic approaches in high-burden settings.

Methodology
The authors conducted a pragmatic, stepped-wedge cluster randomized controlled trial (NCT06017843) embedded within the National TB Control Program (NTP) of Pakistan, supported by the Global Fund and implemented by Mercy Corps.

• Study Design: The trial involved 30 mobile X-ray van teams operating across 68 districts in four provinces. The design utilized a stepped-wedge approach where all teams began in the control arm (routine site selection) and transitioned sequentially to the intervention arm over a 13-month period (August 2023 – September 2024).
• Intervention: The intervention arm utilized MATCH-AI, a decision-support software that employs Bayesian modeling to integrate programmatic surveillance data and contextual variables (e.g., income, population density, immunization coverage) to predict TB risk and identify high-prevalence "hotspots." Teams were provided with GPS coordinates of the highest-risk polygons and instructed to conduct camps at these locations.
• Control: The control arm continued with standard site selection practices, guided by District Field Supervisors using local knowledge, historical notification data, and stakeholder consultation without systematic quantitative targeting.
• Participants: The study included 269,254 individuals screened across 3,936 camps. Children (<12 years) and pregnant women were excluded from chest X-ray screening per national guidelines but included in attendance counts.
• Outcomes: The primary outcome was Camp Positivity Yield (number of bacteriologically confirmed TB cases per camp). Secondary outcomes included Camp Positivity Rate (cases per population screened), All-Forms TB Yield, and All-Forms TB Rate.
• Analysis: The primary analysis was Intention-to-Treat (ITT). A "validated" per-protocol analysis was also conducted, including only intervention camps where GPS verification confirmed the camp was conducted within a 5-km radius of the AI-recommended location. Generalized linear mixed models (negative binomial and Poisson) were used to adjust for clustering and secular trends.

Key Results

• Intention-to-Treat (ITT) Analysis: The ITT analysis, which included all camps conducted by teams after they transitioned to the intervention arm regardless of adherence to specific AI locations, showed a 7% higher Camp Positivity Yield in the intervention group compared to the control. However, this difference was not statistically significant (Adjusted Risk Ratio [RR] 1.07, 95% CI: 0.94–1.22).
• Validated Per-Protocol Analysis: When restricting the analysis to camps that strictly adhered to the AI recommendations (within 5 km of the target), the intervention demonstrated a statistically significant 32% increase in Camp Positivity Yield compared to the control (Adjusted RR 1.32, 95% CI: 1.12–1.54).
• Secondary Outcomes: In the validated per-protocol analysis, the Camp Positivity Rate was also significantly higher in the intervention arm (Adjusted RR 1.20, 95% CI: 1.02–1.41). No significant differences were observed for All-Forms TB outcomes.
• Subgroup Findings:
  • Baseline Yield: The intervention was most effective in districts with moderate baseline yields (0.5%–1%), where the adjusted RR was 1.57 (95% CI: 1.13–2.18). Effects were not statistically significant in low or high baseline yield districts.
  • Geography: The intervention yielded significantly higher results in rural districts (Adjusted RR 1.43, 95% CI: 1.23–1.65) and in the Punjab province (Adjusted RR 1.51, 1.17–1.95).
• Adherence: Adherence to AI recommendations varied significantly by province (highest in Khyber Pakhtunkhwa at 44.5%, lowest in Balochistan at 8.3%), with only 56% of reported intervention camps verified as being within the 5-km target radius.

Significance and Claims
The paper claims to be the first prospective evaluation of a geographical targeting strategy for TB screening in a high-incidence setting. Its primary significance lies in demonstrating that while AI-guided targeting has the potential to substantially improve the efficiency of active case finding, its success is contingent upon strict adherence to the recommended locations.

The authors argue that the divergence between the non-significant ITT results and the significant per-protocol results highlights that deviations from precise AI-recommended sites dilute the intervention's impact. The study provides empirical evidence that:

1. TB distribution is spatially heterogeneous, and targeted screening in predicted hotspots can outperform routine, experience-based site selection.
2. AI tools like MATCH-AI can serve as supportive decision-making instruments to optimize resource allocation in high-burden, low-resource settings without necessarily expanding the total number of screenings.
3. The greatest benefit of such tools may be realized in districts with moderate baseline yields, where routine strategies may miss undetected pockets of transmission.

The authors conclude that integrating spatial intelligence into routine operational planning is a viable strategy to improve early diagnosis and accelerate TB elimination, provided that implementation challenges regarding field adherence and operational constraints are addressed. They do not claim the software replaces human judgment but rather suggests it should be integrated with contextual human knowledge to maximize effectiveness.