Google Research at The Check Up: from healthcare innovation to real-world care settings

Google used its annual Check Up event to lay out a decade's worth of healthcare AI work, and the throughline this year is that these systems are quietly moving out of the lab and into actual clinics, wearables, and public health offices. That's a different story than the usual pile of benchmark announcements.

The clinical results are the ones worth paying attention to. In studies with Imperial College London and the NHS, published in Nature Cancer, Google's breast cancer detection system caught 25% of so-called interval cancers, the ones that usually slip past screening and only show up once a patient has symptoms. Baked into existing radiology workflows, the system also showed it could cut down radiologist workload rather than add to it. Separately, a diabetic retinopathy screening tool built with eye hospitals in India, Thailand, and Australia has now delivered over a million screenings, with diagnoses in as little as two minutes — meaningful when the disease is preventable but only if caught early.

On the consumer side, Google and Fitbit tested something called a Personal Health Agent, essentially a bundle of AI personas acting like a data scientist, a domain expert, and a coach rolled into one, built to outperform single-purpose step-counting apps. It's a bet that people stick with health habits better when the app feels less like a spreadsheet and more like a team paying attention to them. Meanwhile AMIE, a multi-agent system from Google Research and DeepMind, is being piloted with Beth Israel Deaconess Medical Center to handle pre-visit history-taking and flag urgent symptoms, and Google partnered with Included Health on a national telehealth study.

There's also a developer-facing layer: MedGemma, part of Google's open-weight Health AI Developer Foundations kit, is already showing up in the real world — AIIMS in New Delhi is using it for triage and dermatology screening, while Singapore's Ministry of Health is fine-tuning it for primary care. A Kaggle challenge built around MedGemma pulled in more than 850 submissions, with winners due next week.

And the ambitions stretch past individual patients. Google Earth AI's geospatial models were combined with survey data by researchers at Mount Sinai and Boston Children's to map measles vaccination coverage down to the ZIP code, exposing undervaccinated clusters tied to recent outbreaks. On the research side, tools like Co-Scientist, Gemini Deep Think, and a new genomics tool called DeepSomatic are aimed at accelerating hypothesis generation and catching cancer mutations that older tools missed. None of this is finished science, but it's a lot more clinic and a lot less slide deck than most AI health announcements manage.