Can AI Help Find the Missing Patients? Inside the SPOT-TB Trial in Pakistan

Tuberculosis remains one of the world's deadliest infectious diseases, and nowhere is the challenge sharper than in high-burden countries like Pakistan. A newly published study protocol in BMJ Open Respiratory Research describes an ambitious national trial, SPOT-TB, that puts a simple but powerful question to the test: if we let artificial intelligence tell us where to look for TB, do we find more of it?

At the heart of the trial is EPCON's Epi-control platform, referred to in the publication as MATCH-AI. (The naming is historical since Pakistan was one of the first countries where the solution was deployed, and the tool carried the MATCH-AI name there before Epi-control became the platform's broader identity.)

The problem: a case-detection gap of more than 50%

Pakistan has the fifth-highest TB burden in the world, with an estimated 573,000 people developing TB each year and around 42,000 deaths. Yet in 2020, only about 48% of those cases were actually diagnosed. That means more than half of the people who fell ill were never formally found by the health system, continuing to suffer, and potentially continuing to transmit the disease.

Closing that gap is a top priority for Pakistan's National TB Programme.

Why active case finding alone hasn't been enough

One of the main strategies to catch undiagnosed cases is active case finding (ACF) — going out into communities and screening people rather than waiting for them to show up sick at a clinic. Pakistan has invested heavily here, rolling out mobile X-ray vans equipped with computer-aided detection (CAD) and rapid molecular testing (Xpert MTB/RIF), largely through Mercy Corps and partners with Global Fund support.

The catch: these "chest camps" are resource- and labour-intensive, and their yields have often come in lower than hoped. A recent Joint Programme Review Mission flagged exactly this issue and recommended a smarter, more geographically targeted approach.

TB doesn't spread evenly

The rationale for targeting rests on a well-established fact: TB is not spread uniformly across a population. Like many infectious diseases, it clusters in "hotspots" — specific neighbourhoods and communities where prevalence runs far higher than average. A review of 168 studies found spatial heterogeneity in every single one, and a Karachi study of nearly 200,000 people screened over two years confirmed the same pattern locally.

If TB is concentrated, then screening everywhere equally is inefficient. The smarter move is to send limited resources to where the disease actually is.

Enter Epi-control platform (MATCH-AI)

This is where EPCON's platform comes in. Developed together with the KIT Royal Tropical Institute, Epi-control uses a Bayesian modelling approach to predict TB prevalence down to small "polygons" of roughly 10,000 people.

Crucially, it doesn't rely on any single data source. The model pulls together historical TB notification data, previous ACF results, and a wide range of contextual factors, demographics, income, population density, health indicators such as vaccination coverage, and even climate variables, to estimate where TB is most likely hiding. A dashboard then lets field teams see these predictions on a map to guide where camps should be set up. The system is self-learning: as new data flows in, its predictions sharpen.

How the trial works: a stepped wedge design

SPOT-TB is a pragmatic, stepped wedge cluster randomised controlled trial, a mouthful, but the logic is elegant.

Thirty mobile X-ray teams ("van-teams") working across 72 districts in all four provinces of Pakistan take part. At the start, every team uses the conventional approach to choosing camp sites: local knowledge, field experience, and historical records. Then, month by month, groups of teams are randomly switched over to using Epi-control's platform predictions instead, until, by the end of the 12-month trial, all 30 teams are using it.

Because each team acts as its own control before switching, and because the rollout is randomised, the design offers a rigorous way to isolate the platform's effect while staying fully embedded in the programme's real-world operations. Nothing else about the screening, diagnosis, or treatment process changes.

What success looks like

The primary outcome is straightforward: Camp Positivity Yield, the number of bacteriologically confirmed TB cases found per camp. The trial is powered to detect a 32% increase in cases found in the AI-guided arm compared with conventional site selection.

Why it matters

To the authors' knowledge, SPOT-TB is the first prospective trial to test geographically targeted ACF in a high-burden setting. Until now, the case for targeting has rested mostly on theory and modelling and the only comparable field studies came from the low-burden United States.

If the approach works, the implications reach well beyond Pakistan. It would offer hard evidence that AI-guided targeting can make active case finding more effective and more cost-efficient, and the platform behind it can be deployed in other high-burden settings striving toward TB elimination. In a field where every undiagnosed case is both a personal tragedy and a source of onward transmission, finding people faster and more precisely could genuinely move the needle.

Read the full publication: Geographical targeting of active case finding for tuberculosis in Pakistan using hotspots identified by artificial intelligence software (SPOT-TB), BMJ Open Respiratory Research 2024