Skip to content
← All research

News, TB screening · AI advice · Regulation · Health apps

AI Is Now Telling TB Vans Where to Park. Did It Work?

Pakistan tested AI that picks where X-ray vans set up TB screening. Where teams followed it, they found 32% more TB. Also this week: drug ads that sway AI chatbots, Singapore fast-tracking new treatments, and a pregnancy app that changed nothing.

19 November 20266 min read

Pakistan just finished a large real-world test of AI in public health, covering all four of its provinces. The software didn’t diagnose anyone. It made a simpler decision: where the X-ray van should go next. Also this week: drug ads that sway AI chatbots’ advice, Singapore speeding up access to new treatments, a pregnancy app that didn’t move the needle, and why your smartwatch only helps if you actually wear it.

269,254people screened for TB in Pakistan
3,936mobile screening camps in 68 districts
+32%more TB found where teams followed the AI
No effecton weight from a year-long pregnancy coaching app

The story: AI that chooses the spot, not the diagnosis

Tuberculosis is still one of the world’s deadliest infections, and many people who have it never get tested. In Pakistan, teams drive 30 vans fitted with X-ray machines into towns and villages and set up pop-up screening camps. Each van can only be in one place a day, so where it parks decides how many cases get found.

Normally, teams choose sites from experience and old records. In the SPOT-TB trial, published on 12 November in The Lancet Global Health, teams were switched one by one to software called MATCH-AI, which maps where undiagnosed TB is most likely and recommends locations.

Between August 2023 and September 2024 the vans ran 3,936 camps across all four provinces.

What they found

  • Across every camp, the gain was small (7%) and not statistically significant. It could have happened by chance.
  • But in camps set up within 5 km of where the AI pointed, teams found 32% more confirmed TB per camp. That result was strong.
  • The biggest gains were in rural districts, where a van’s limited days matter most.
Why it mattersThe AI didn't need new machines or new staff. It made an existing, scarce resource, a van with an X-ray, go further. For rural health systems, that kind of tool is often worth more than a flashier diagnostic.

The catch

The honest headline is “promising, not proven”. The strong result only counts camps where teams followed the software, and teams that followed it may differ in other ways. The researchers also say routine programme data couldn’t show exactly how well each team carried out the plan. And the study didn’t report cost per case found, which is what a health ministry needs before rolling it out nationwide.

There’s a fairness question too: software trained on where TB was found before could keep sending vans back to the same places, missing communities that have never been screened.

Also this week

Drug ads can sway what AI chatbots recommend

More people now ask AI chatbots about their health. A study from the Icahn School of Medicine at Mount Sinai in New York, published in The Lancet Digital Health, tested what happens when a clearly labelled drug advert appears just before the question. The team ran 12 AI models, including ones from Google, OpenAI and Anthropic, through 33 doctor-written patient cases, collecting over 258,000 answers.

  • In 4 of the 33 cases, the models went from never choosing the advertised drug to choosing it every time.
  • The models repeated the ad’s claims more often than they admitted it. They mentioned the ad in only about 30% of the written answers checked.
  • The good news: the ads didn’t clearly push the AI towards unproven supplements or clearly worse drugs. The study also didn’t test whether real prescriptions changed.
Why it mattersIf chatbots start carrying ads, the advice can tilt without saying so. The researchers want AI makers to test for this, label any advertising, and keep it apart from medical advice.

Singapore speeds up the path from approval to subsidy

In Singapore, a new treatment is first approved as safe by the Health Sciences Authority (HSA). Then the Agency for Care Effectiveness (ACE) decides whether it’s good value for public funding. Until now those steps ran one after the other. From this month, a new priority track runs them at the same time for selected innovative products.

~1 yearuntil the first funding decision after approval, today
6–9 monthswith the new parallel track
30 → 15 dayspossible approval time for some lower-risk late-stage trials

HSA says safety and effectiveness standards stay the same, and the two decisions stay independent. Since 1 November, HSA has also taken over the Ministry of Health’s regulatory functions, putting more of the rules under one roof.

A pregnancy coaching app made no difference to weight

The Bump2Baby and Me trial (The Lancet Digital Health, 12 November) gave 865 pregnant people at higher risk of gestational diabetes, in Australia, Ireland, Spain and the UK, personalised smartphone coaching from pregnancy until their baby’s first birthday.

−0.04BMI change with the app (kg/m²)
+0.04BMI change with usual care
18% vs 25%adverse events, app vs usual care; none linked to the app

A year after birth, weight change was the same in both groups. The app was safe, and the researchers suggest future programmes aim at a wider range of mothers’ health, not weight alone. Trials that publish a “no effect” result like this one are useful: they stop money going into tools that don’t work.

Your smartwatch only helps if you keep wearing it

A viewpoint in The Lancet Digital Health (17 November), by researchers including teams in London and Singapore, argues that wearables are let down by people taking them off. The authors list nine factors that affect how long people wear a device, from comfort to how relevant it feels.

Their point: not every condition needs 24/7 tracking. Catching an irregular heartbeat that comes and goes needs constant monitoring. Keeping an eye on stable blood pressure may only need regular check-ins. Designing for the condition makes people more likely to stick with it. This is an expert opinion piece, not a trial.

What this means for you

  • AI in healthcare isn’t only about diagnosis. Some of the most useful tools quietly improve logistics: where to screen, who to call first.
  • “It works” depends on people using it as designed. A good algorithm that teams ignore changes nothing.
  • Look for trial results before trusting a health app. Including the disappointing ones.
  • Treat AI health advice as a starting point. Check it with a doctor or pharmacist, especially if it points you to a brand-name drug.
  • If you wear a health tracker, ask what it’s for. The right amount of wear depends on what you’re tracking.

Sources

  1. Mahfooz A, et al. SPOT-TB trial. The Lancet Global Health, 12 Nov 2026. doi.org/10.1016/j.langlo.2026.104090
  2. O’Reilly SL, et al. Bump2Baby and Me trial. The Lancet Digital Health, 12 Nov 2026. doi.org/10.1016/j.landig.2026.101059
  3. Lu L, et al. Wear time matters (Viewpoint). The Lancet Digital Health, 17 Nov 2026. doi.org/10.1016/j.landig.2026.101044
  4. Mount Sinai. “Ad-verse Effects: Pharmaceutical Advertising and Drug Recommendations by Consumer-Facing AI in a Controlled Study.” The Lancet Digital Health, 11 Nov 2026. News release (EurekAlert)
  5. Health Sciences Authority, Singapore. Strengthening local and global partnerships to accelerate healthcare innovations. hsa.gov.sg

Updated 19 November 2026 with the drug-ad AI study and Singapore’s new approval track. Trial figures are from each paper’s published abstract (PubMed records 43020100, 43020746 and 43045610); the drug-ad figures are from Mount Sinai’s news release; the Singapore figures are from HSA’s announcement. This article is for general information and isn’t medical advice.