Four kilometers underground, there is no cloud. Google's TB screening AI took seven years to get there

About four kilometers below the surface at the Mponeng gold mine in South Africa, miners are now screened for tuberculosis on site. A portable digital X-ray kit takes the image, an AI model reads it without any internet connection, and anyone flagged is referred for a confirmatory molecular test during the same visit.

The model behind it was first described in a research paper in 2019. Google Research published the story of what happened in between on October 23, and it is one of the clearest accounts available of how long the road from a validated model to routine clinical use actually is.

The timeline is the lesson

From paper to the mine floor

20192026

1. ResearchModel matches radiologists on active TB (Radiology)
2. Independent evaluationStop TB Partnership ranks it among top algorithms; WHO recommends computer-aided detection
3. Prospective validationMulti-site study (NEJM AI)
4. CertificationCE MDR Class IIb and ISO 13485 via Nexus AI
5. DeploymentAbout 40 facilities, offline in a mine

In 2019, a study in Radiology showed the deep learning model matched experienced radiologists in detecting active pulmonary TB. In 2021, an independent evaluation by the Stop TB Partnership, published in The Lancet Digital Health, ranked it among the top-performing screening algorithms, and the World Health Organization recommended computer-aided detection for TB screening and triage that year. In 2024, a prospective multi-site validation in NEJM AI confirmed high sensitivity across diverse populations.

None of that made it a medical device. That step came through Nexus AI, a medtech company in Pretoria that turned Google’s algorithms into a certified clinical tool: CE MDR Class IIb under the EU Medical Device Regulation and ISO 13485 for its quality management system. WHO prequalification is being pursued through an agreement with FIND, and there is an “ongoing pathway” toward U.S. FDA clearance. Neither has been granted.

What deployment looks like

Google reports about 40 healthcare facilities across six countries, with current use in South Africa, Sierra Leone and Côte d’Ivoire and expansion underway in the Democratic Republic of the Congo, Zambia and Cameroon. More than 90 clinicians have evaluated over 25,000 chest radiographs, and results arrive in as little as 45 seconds instead of weeks. More than 3,000 free screenings went to under-resourced public clinics, toward a goal of 10,000 a year.

Two design facts matter more than the counts. First, screening is not diagnosis: the model triages, flagged patients go to a confirmatory test, and clinicians make the final decisions. Second, the system runs offline. In a mine, a rural clinic or anywhere with intermittent connectivity, a cloud-only model would simply not be there when needed. We have argued that sovereign AI is defined by what keeps working when the connection is cut; this is the clinical version of that test.

The next chapter: open models, national data

The newer work uses Google’s open medical models. MedGemma, available in 4B and 27B versions, and MedSigLIP, a lightweight vision-language model, can be fine-tuned with small local datasets. Indonesia’s Ministry of Health is fine-tuning both on domestic chest radiographs and running them on local infrastructure, so patient data stays within the country, as it integrates TB screening into its national program. Its target of screening 50 million people a year is an ambition, not a result, and the post presents it that way.

This is a different model of ownership from the Nexus AI path. A certified product puts a manufacturer on the hook for performance. A ministry fine-tuning open weights takes on the validation, monitoring and regulatory obligations itself.

What transfers to any regulated AI program

For organizations building AI for regulated decisions, in health or elsewhere, the account is useful for what it does not skip:

1. Evidence comes in layers. A retrospective benchmark, an independent evaluation and a prospective study answer different questions. Procurement and regulators will ask for all three.
2. Certification is a separate project. Turning a model into a device took a company with a quality system, and the certification is jurisdiction-specific: CE marking does not grant FDA clearance or WHO prequalification.
3. The decision boundary must be explicit. “Triage, then confirm” is a design, not a disclaimer, and it is what made the deployment acceptable.
4. The environment sets the architecture. Offline operation was a requirement discovered on the ground, not a feature chosen in a lab.

The model was ready in 2019. The system that could be used four kilometers underground took most of the following seven years, and almost all of that work was outside the model.