Three medical AI announcements came out on November 6. None of them is a product launch. Each is about testing AI against something hard to fake, like patient records, lab results or hospital outcomes. That is where the field is heading after two years of chatbot demos.
Viz.ai finds heart disease the records missed
Two abstracts on Viz HCM, the AI-ECG algorithm in Viz.ai’s Cardio Suite, were accepted for the Hypertrophic Cardiomyopathy Medical Society meeting held alongside the American Heart Association Scientific Sessions in Chicago. Viz HCM reads standard 12-lead ECGs and was the first FDA-cleared AI tool for spotting hypertrophic cardiomyopathy. It was built under a multi-year deal with Bristol Myers Squibb, which sells a drug for the condition.
The University of Colorado ran the algorithm over ECGs from more than 29,000 patients. Close to 40 percent of the patients it flagged had no matching diagnosis code in their records. At Beth Israel Deaconess, patients flagged by the model had a 2.68 times higher hazard of adverse events, and the model caught suspected cases that standard ECG measurements did not.
The appeal is that hospitals already hold millions of ECGs. Run software over the archive and you get a list of people to call back. The catch is the funding. The study was supported by a BMS company, and every new diagnosis is a potential new patient for its drug. That does not make the data wrong, but it is why independent replication matters.
Cancer Grand Challenges puts AI co-scientists to work
Cancer Grand Challenges named the first three teams for AURORA, a 10 million dollar program testing agentic AI co-scientists that analyse data, suggest hypotheses and rank experiments while humans keep the final call. Cancer Research UK put 6 million dollars into the core AI system in March, built in Marinka Zitnik’s lab at Harvard Medical School. Cancer Grand Challenges is giving about 4 million to three pioneer teams.
NexTGen, led by Nicky McGranahan at UCL, works on next-generation CAR-T therapy, including for children with solid tumours. PROMINENT, led by Nuria Lopez-Bigas at IRB Barcelona and Emma Lundberg at Stanford, studies how exposures push normal cells toward cancer. MATCHMAKERS, led by Sergio Quezada and Welles Robinson at UCL with Peter Bruno at UC San Francisco, uses interacting AI agents to find tumour antigens and the T-cell states most likely to attack them. Their lab results feed back into the shared system. That loop is the real test. An AI that proposes experiments is easy to build. One whose proposals beat a good postdoc’s is the claim to check.
Feinstein targets delirium with wearables
Northwell Health’s Feinstein Institutes got a 3.4 million dollar NIH grant to build machine learning models for delirium. Clinicians miss it in more than 75 percent of cases, and up to 40 percent of admitted patients may develop it. The team, led by Liron Sinvani, will combine health records with continuous wearable data from what Northwell calls the largest in-hospital wearable dataset in the US. The goal is to flag patients who arrive delirious and predict who is at risk, then build the tools into ward routines with nurses and doctors.
Workflow is the part that decides whether this helps. A risk score nobody acts on is just one more alert in a hospital full of them.
The common thread
Each project measures AI against outcomes, not benchmarks. That is slower and less flashy than a model launch. It is also the only way medical AI earns a place in routine care. Watch for follow-up numbers. How many flagged HCM patients were confirmed on echo, how many AURORA hypotheses held up at the bench, and whether delirium rates fall on wards that use the Feinstein tools.