On October 23, 2026, Google Research published "Bringing foundational models to the front lines of TB screening". It follows the October 21 post on blind-sweep ultrasound, which this digest covered on October 22, and has the same theme: research models only matter once they run where specialists are scarce. The post puts the bottleneck plainly: "radiologists are scarce, and imaging infrastructure is often days away."
From study to certified device
Google traces the work through a 2019 benchmark, a 2021 evaluation with the Stop TB Partnership in The Lancet Digital Health, and a prospective multi-site validation published in NEJM AI in 2024, which the post says showed screening "maintains high sensitivity (minimizing missed cases) across diverse patient populations". The deployment partner is Nexus AI, a Pretoria medtech company. Under the EU Medical Device Regulation, "Nexus AI achieved Class IIb certification". Its device "screens chest X-rays for lung diseases like TB in as little as 45 seconds", and "over 90 clinicians have evaluated more than 25,000 chest radiographs", with clinicians making the final decision.
The tool runs in "approximately 40 healthcare facilities" in South Africa, Sierra Leone and Cote d'Ivoire, with expansions under way in the Democratic Republic of the Congo, Zambia and Cameroon. The post describes the six together as "six countries", but three are expansions in progress. The memorable deployment is at a gold mine, where Nexus AI put "ultra-portable digital X-ray kits and offline AI four kilometers underground" with same-visit referral for molecular testing. Nexus AI has delivered more than 3,000 free screenings and is working with FIND toward WHO Prequalification, which global procurement agencies require. FDA clearance is described as an "ongoing pathway".
The Indonesia model
The more transferable part is Indonesia, which Google says has "the world's second-highest TB burden" and aims at "screening 50 million citizens annually". The post notes that "Reviewing 50 million X-rays manually is infeasible, and proprietary cloud software poses data residency challenges." So "the Ministry is fine-tuning MedGemma and MedSigLIP on domestic chest radiographs", and "Because the models run on local infrastructure, patient data remains within national borders". MedSigLIP is a lightweight vision-language model that can classify images with plain-language labels or be fine-tuned "with only a small set of local examples". Google says MedGemma has passed 10 million downloads.
Cough as a first filter
Where no X-ray is available, Google points to HeAR, an acoustic model "Trained on over 300 million audio segments" that screens from cough sounds on a phone microphone. CIDRZ in Zambia, the Aurum Institute in Ghana and the HELINA network are "building the clinical evidence for cough-based triage" -- in other words, it is not yet validated for that use.
What remains uncertain. The post gives no accuracy figures for the newer deployments, no outcome data such as additional cases found or treatment started, and no independent evaluation of the Indonesian fine-tuned models. WHO Prequalification and FDA clearance are pending.
Analysis: India, which carries the largest share of the world's TB cases in WHO's global reporting, faces the same arithmetic as Indonesia: tens of millions of images and too few radiologists. The pattern to copy is not a vendor product but an architecture: open medical foundation models, fine-tuned on national data, served on national infrastructure, with a certified device for the front line and a human signing off. For health-AI teams, the operational lessons are offline inference for sites without connectivity, a referral step in the same visit so a positive screen is not lost, and a clear line between a validated tool (X-ray triage) and a promising one (cough triage). Procurement teams should ask for outcome data, not just sensitivity.
Sources and evidence
Google Research (October 23) says chest X-ray AI from its work now runs as Nexus AI's CE Class IIb device in about 40 facilities across South Africa, Sierra Leone and Cote d'Ivoire, including offline four kilometres underground, while Indonesia fine-tunes the open MedGemma and MedSigLIP models on domestic X-rays to screen 50 million people a year with data kept in-country; cough-based HeAR triage is still gathering clinical evidence and no new accuracy or outcome figures were published.