This week: Google Research shared how its tuberculosis X-ray AI went from a 2019 paper to a certified tool in about 40 clinics across Africa, plus one very unusual site: a gold mine four kilometers underground.

Elsewhere: What "presumptive TB" actually means, why open models matter for a country screening 50 million people a year, and an AI that listens to coughs.

The New Prime Signal

Most AI health stories end with a promising paper and a line about "future work." This one has a sequel, and the sequel involves a mine shaft. In a new Google Research post published October 23, the team behind Google's tuberculosis screening work laid out the whole journey: research benchmark, clinical trials, regulatory paperwork, and finally a tool that screens chest X-rays "in as little as 45 seconds." It now runs in clinics in Africa and, thanks to portable X-ray kits and offline AI, four kilometers underground at one of the deepest mines in the world. It's not every day a machine learning model gets a hard hat.

What happened: Google partnered with Nexus AI, a medical technology company based in Pretoria, South Africa, which turned Google's chest X-ray algorithms into a certified clinical product. According to Google, the tool earned CE MDR Class IIb certification under EU medical device rules, and Nexus AI is pursuing World Health Organization prequalification, with a pathway toward U.S. FDA clearance still in progress. Google says the tool operates in "approximately 40 healthcare facilities" in South Africa, Sierra Leone, and Côte d'Ivoire, with expansions underway in the Democratic Republic of the Congo, Zambia, and Cameroon. So far, "over 90 clinicians have evaluated more than 25,000 chest radiographs," and Nexus AI has delivered more than 3,000 free screenings to under-resourced public clinics. Google shared the update this week at IDWeek 2026, an infectious disease conference in Washington, D.C.

Why it matters: TB is preventable and curable, yet Google's post puts the toll at more than 1.23 million deaths a year. A big bottleneck is people, not pills: reading chest X-rays for subtle signs of TB takes a trained radiologist, and in many high-burden areas the nearest one is days away. Google says the tool cuts turnaround "from weeks to under 45 seconds." The mine shows why that matters. Deep-shaft miners breathe silica dust and face higher lung disease risk. At the Mponeng Gold Mine in South Africa, miners are now screened on-site, and anyone flagged can be sent straight for a confirmatory lab test. Instead of weeks of back-and-forth, the whole process can happen in one visit.

The fine print: First, this is a screening tool, not a diagnosis. Google describes it as identifying "people with presumptive TB for follow-up testing," and says clinicians still make the final call. A flagged X-ray means "go get the real test," not "you have TB." Second, the numbers in this post come from Google and its partner. The accuracy research behind it has been published over the years, in Radiology (2019), The Lancet Digital Health (2021, an independent Stop TB Partnership evaluation), and NEJM AI (2024), but the deployment figures (40 facilities, 25,000 X-rays) are self-reported, and the post doesn't share new accuracy results from the field. Third, "45 seconds" is the time to screen an X-ray, not to cure anyone, and three of the six countries are still listed as expansions. And FDA clearance and WHO prequalification are both still in progress.

This is the unglamorous, good kind of AI news. No benchmark crown, no chatbot drama, just a long grind from paper to product in places where a radiologist is hard to find. The most interesting move is the shift from one partner's product to open models anyone can adapt, which is how a tool built for a few dozen clinics could end up helping a whole country. We'd still like to see independent numbers from the field. But "an X-ray machine and a laptop at the bottom of a gold mine" is a sentence we're happy to have written.

One company's product can only reach so many clinics. Google's bigger bet is giving health systems the parts so they can build their own.

How it works: Through its Health AI Developer Foundations program, Google has open-sourced chest X-ray embeddings and MedSigLIP, a lightweight model that matches medical images with the words doctors use to describe them. That means a local team can sort or search X-rays with plain-language labels, or fine-tune the model with "only a small set of local examples." On top of that sits MedGemma, a family of open-weight models in 4B and 27B sizes that can combine X-ray findings with a patient's history, answer clinicians' questions, and run on a local device without internet. Google says MedGemma has passed "10 million downloads." Indonesia, which has the world's second-highest TB burden, is using this route. Its health ministry aims to screen 50 million people a year and is fine-tuning MedGemma and MedSigLIP on its own X-rays, on its own servers, so patient data stays inside the country.

Try it: If you're a developer or researcher, MedGemma and MedSigLIP are available through Google's Health AI Developer Foundations site, along with documentation and terms of use. These are research building blocks, not doctors, so anything built for patient care needs its own testing and approval. Google also highlighted HeAR, a model trained on "over 300 million audio segments" that screens for TB from cough sounds recorded on an ordinary smartphone. It's still at the research stage, with groups in Zambia and Ghana building the clinical evidence. If it pans out, a phone could become the first step in the screening chain, before anyone needs an X-ray at all.