Google Research details how its tuberculosis screening AI reached clinics through Nexus AI, MedGemma and HeAR

Google Research has published a detailed account of how its chest X-ray models for tuberculosis (TB) screening moved from academic benchmarks into certified clinical use, and how the company now intends to scale that work through open health foundation models rather than a single proprietary tool. The post, dated October 23, 2026 and written by Preeti Singh, Lead Technical Program Manager, and Shravya Shetty, Distinguished Software Engineer, was shared alongside Google's presence at IDWeek 2026 in Washington, D.C.

The account brings together three threads that had previously been reported separately. The first is the clinical translation of Google's TB detection algorithms by Nexus AI, a medtech company based in Pretoria, which has obtained CE MDR Class IIb certification for a diagnostic tool built on them. The second is Health AI Developer Foundations (HAI-DEF), the family of open models that includes chest X-ray embeddings, the MedSigLIP vision-language encoder and the MedGemma open-weight multimodal models. The third is HeAR (Health Acoustic Representations), an audio foundation model that research partners are using to study cough-based TB triage from ordinary smartphone recordings.

The figures Google reports are operational rather than benchmark-driven. According to the post, the Nexus AI tool runs in about 40 healthcare facilities across six countries, more than 90 clinicians have evaluated over 25,000 chest radiographs with it, and the reading time per radiograph is under 45 seconds, compared with turnaround times that Google describes as weeks in the settings where the tool is deployed. Indonesia's Ministry of Health is separately fine-tuning MedGemma and MedSigLIP on domestic chest X-rays as part of a national screening ambition of 50 million people per year.

··········

THE CLINICAL PROBLEM AND THE EVIDENCE BEHIND THE CHEST X-RAY MODEL.

Google frames the work around a well-documented bottleneck. Tuberculosis causes an estimated 1.23 million deaths a year, and many of the regions with the highest burden have very few radiologists available to read chest radiographs, which remain one of the main screening and triage instruments for the disease. When an X-ray can be taken locally but has to wait days or weeks for an expert reading, patients may be lost to follow-up before a diagnosis is confirmed and treatment starts.

The post traces the evidence for Google's chest X-ray approach through three published milestones. A 2019 benchmark study in Radiology reported that deep learning could detect active pulmonary TB at a level matching experienced radiologists. In August 2021, an independent evaluation by the Stop TB Partnership, published in The Lancet Digital Health, ranked Google's model among the top-performing TB screening algorithms it tested. The same year, the World Health Organization recommended computer-aided detection (CAD) for TB screening and triage, which created a policy basis for using such software in national programs.

The third milestone was a prospective, multi-site validation published in NEJM AI in September 2024. Google says this study reported high sensitivity across diverse populations and also showed that the system could flag incidental pulmonary and cardiac abnormalities beyond TB. The blog post does not restate the sensitivity and specificity values from these papers, so the exact operating points of the deployed system cannot be derived from the announcement itself.

··········

HOW NEXUS AI TURNED THE ALGORITHMS INTO A CERTIFIED DEVICE.

The central claim of the post is that research performance alone does not put a model into a clinic. Google describes this as a "translation gap" and says it addressed it through a partnership with Nexus AI, which packaged the algorithms into a diagnostic product and took it through regulatory and quality processes.

Google lists four regulatory and quality steps for the Nexus AI tool. The product has CE MDR Class IIb certification under the EU Medical Device Regulation 2017/745, which Google says was based on prospective clinical evaluations. Nexus AI also operates under ISO 13485, the quality management standard for medical device manufacturers. The company has signed an agreement with FIND, the global diagnostics alliance, to pursue WHO prequalification, and a pathway toward clearance by the U.S. Food and Drug Administration is described as ongoing rather than completed.

........

Step	Status reported by Google
CE MDR Class IIb (EU MDR 2017/745)	Certified
ISO 13485 quality management	Compliant
WHO prequalification	Agreement with FIND to pursue it
U.S. FDA clearance	Pathway ongoing

........

According to the post, the software analyzes a chest radiograph in as little as 45 seconds. Its interface highlights findings that are inconsistent with normal lung fields, including findings suggestive of TB, and flags cases that need further investigation. That design places the tool in a triage role: it prioritizes which patients should go on to confirmatory testing rather than issuing a standalone diagnosis.

··········

DEPLOYMENT FIGURES AND THE UNDERGROUND SCREENING PROGRAM.

The deployment numbers in the post describe a footprint that is still modest in absolute terms but spread across several health systems. Three countries are listed as active, South Africa, Sierra Leone and Côte d'Ivoire, while expansions are described as underway in the Democratic Republic of the Congo, Zambia and Cameroon. Google counts all six when it gives the country total.

........

Metric	Figure reported by Google
Countries	6 (3 active, 3 expanding)
Healthcare facilities	About 40
Clinicians using the tool	More than 90
Chest radiographs evaluated	More than 25,000
Reading time per radiograph	Under 45 seconds
Free screenings delivered	More than 3,000
Sponsored screening target	10,000 per year

........

The most unusual deployment described is at the Mponeng Gold Mine in South Africa, where Nexus AI worked with the Stop TB Partnership, Harmony Gold Mine and the South African Department of Health. Ultra-portable X-ray kits and AI that runs offline were installed about four kilometers underground. Google says this allows screening, triage and confirmatory molecular testing to happen during a single visit, which removes the delay between an abnormal X-ray and a laboratory result for a workforce with elevated TB exposure.

The offline requirement in this setting is a technical detail with practical weight. A screening model that depends on a cloud connection cannot work reliably in a mine shaft or in many rural clinics, so the deployment depends on inference running on local hardware next to the X-ray unit.

··········

FROM A SINGLE TOOL TO OPEN FOUNDATION MODELS.

The second half of the post explains why Google is shifting emphasis from one certified product to open models that other organizations can adapt. Through HAI-DEF, Google has released chest X-ray embeddings and MedSigLIP, and it has published MedGemma as open-weight multimodal models. The post says MedGemma comes in 4B and 27B configurations, can run locally without internet access, and has passed 10 million downloads.

The HAI-DEF documentation gives more detail on the components. MedSigLIP is a two-tower encoder based on SigLIP-400M, with a vision encoder and a text encoder of roughly 400 million parameters each, taking 448 x 448 images and text of up to 64 tokens. It was trained on chest X-rays, dermatology, ophthalmology, histopathology and CT and MRI slices paired with reports, and Google's model card reports a zero-shot average AUC of 0.844 across 13 chest X-ray findings. The same encoder handles image interpretation inside MedGemma. MedGemma itself is built on Gemma 3, with a 4B multimodal model in version 1.5 and 4B multimodal, 27B text-only and 27B multimodal variants in version 1.

........

Component	Type	Key specifications from Google documentation
MedSigLIP	Image-text encoder	About 400M + 400M parameters, 448 x 448 input
MedGemma	Generative multimodal model	4B and 27B, built on Gemma 3, open weights
HeAR	Audio encoder	ViT-L masked autoencoder, 2-second 16 kHz clips
Chest X-ray embeddings	Image embeddings	Released through HAI-DEF

........

Google's documentation is explicit about the limits of these releases. MedGemma is described as a developer model that "isn't yet clinical-grade" and requires validation for each intended use, and MedSigLIP is described as providing no medical functionality on its own. All HAI-DEF models are distributed under the Health AI Developer Foundations terms of use rather than a standard open-source license, so "open" here means downloadable weights under Google's own terms, not open source in the OSI sense.

··········

INDONESIA AND THE CASE FOR SOVEREIGN FINE-TUNING.

The clearest example of the foundation-model strategy is Indonesia. Google says the country's Ministry of Health aims to screen 50 million citizens a year for TB and is fine-tuning MedGemma and MedSigLIP on domestic chest radiographs. The adapted models run on local infrastructure so that patient data stays within national borders.

This arrangement illustrates why open weights matter in health care specifically. A ministry that cannot send radiographs to a foreign cloud service for legal or political reasons can still use a model trained by Google, provided it can download the weights, adapt them to local imaging equipment and population characteristics, and validate them on its own data. The post does not give accuracy figures for the Indonesian fine-tuned models or a deployment timeline, so the initiative should be read as an ongoing program rather than a completed rollout.

··········

HEAR AND COUGH-BASED TRIAGE.

The third component is HeAR, which Google describes as trained on more than 300 million audio segments. The model card is more precise: about 313.3 million two-second clips, roughly 174,000 hours of audio, drawn from public non-copyrighted video and filtered with a health acoustic event detector. HeAR is a masked autoencoder with a ViT-L architecture that takes two-second, 16 kHz mono clips and outputs a 512-dimensional embedding per clip. It produces representations, not diagnoses, so any TB classifier has to be trained and validated on top of it.

Google lists three research collaborations working on cough-based TB screening with standard smartphone microphones. The Centre for Infectious Disease Research in Zambia (CIDRZ) is one partner. The Aurum Institute is working in Ghana with a $250,000 grant from Google.org, and HELINA leads a regional collaborative funded with $1.5 million. Google presents acoustic triage, mobile X-ray screening and multimodal clinical reasoning as a cascade: HeAR to decide who should be imaged, HAI-DEF models to read the image, and MedGemma to combine imaging findings with clinical history.

The model card also lists limitations relevant to this use, including the focus on short clips, possible bias related to demographics and recording devices, and a model size that Google says is currently too large for on-device deployment.

··········

WHAT IS ESTABLISHED AND WHAT REMAINS OPEN.

The verified elements of the announcement are the regulatory status of the Nexus AI product in Europe, the deployment counts Google reports, the specifications of the HAI-DEF models in their public documentation, and the funding amounts attached to the HeAR research partnerships. The peer-reviewed record behind the chest X-ray model, from the 2019 Radiology study to the 2024 NEJM AI prospective validation, is cited by Google but not quantified in the post.

Several important questions are left open. WHO prequalification and FDA clearance for the Nexus AI tool are still in progress, the Indonesian fine-tuning effort has no published performance data, and cough-based triage with HeAR remains at the stage of building clinical evidence. The scale of the current deployment, 25,000 radiographs across about 40 facilities, is meaningful as proof of operation but small relative to national screening programs, and the post does not report outcome measures such as additional cases detected or time to treatment.

The broader signal is a change in how Google positions its health AI work: the company is presenting a certified point solution as the first step and open, locally adaptable foundation models as the mechanism for scale, with partners and governments carrying responsibility for validation and regulatory approval in their own settings.

··········

FOLLOW US FOR MORE.

·····

DATA STUDIOS

·····

datastudios.org