|
3, 2, 1: Health AI Brief
Every Friday
June 5, 2026
|
|
AI is reshaping healthcare fast. Below are 3 key AI developments, 2 studies, and 1 takeaway for this week to help you better lead with AI. Target read time: 5 minutes. |
|
3
Market Signals
On June 2, Mayo Clinic and Microsoft announced a collaboration to build a frontier AI model purpose‑built for healthcare. It will be trained on Mayo's de‑identified clinical data and longitudinal patient records, and Mayo will own the model, with access offered to others through Microsoft's Azure AI Foundry. Microsoft AI CEO Mustafa Suleyman said "frontier medical intelligence is around the corner." No financial terms were disclosed. So what?
The interesting part is the ownership. Mayo isn't licensing someone else's AI; it's building its own on its own data and offering it back to the market. I'm expecting more large health systems to start treating their clinical data as an even more valuable model‑building asset, and more companies like Microsoft will be one of many bidders. Read the Microsoft announcement → | Read the Fierce Healthcare coverage → On June 1, the Joint Commission launched the Responsible Use of AI in Healthcare (RUAIH) certification, its first credential covering how provider organizations deploy AI. Developed with the Coalition for Health AI (CHAI), it's voluntary and organized around 5 areas: governance, data management, bias and risk reduction, safety monitoring and validation, and transparency and training. It certifies an organization's use of AI rather than individual AI products, and any health system can apply, accredited by the Joint Commission or not. So what?
Most AI oversight so far has been about regulating the tools: FDA clearances, model cards, device pathways. This now certifies the user, not the tool. The burden of vetting AI is shifting onto the health systems deploying the AI. Read the Joint Commission announcement → | Read the Becker's coverage → On May 29, Massive Bio announced a partnership with TOUCH, The Black Breast Cancer Alliance, and Navigating Trials to embed its AI‑powered clinical‑trial matching directly inside the advocacy groups' patient platforms, at no cost to patients. The company has facilitated more than 19,000 trial matches to date, and a study in ESMO Real World Data and Digital Oncology found its AI matched patients about 4x faster across 3,804 real‑world cancer patients. TOUCH's WhenWeTrial.org platform has had 35,000 patients search for a trial. So what?
Clinical trials have a well-documented diversity problem: the patients enrolled often don't look like the patients who will eventually take the drug. Putting the matching tool where underrepresented patients already are, instead of waiting for them to find it, is a deliberate distribution choice. I'm hoping this is the start of aiming AI at access gaps, not just efficiency gains. |
|
2
Research Studies
A diagnostic study in JAMA Dermatology compared 3 AI models against 652 physicians reading 1,117 real-world skin-lesion cases with clinical images and metadata. Expert dermatologists with over 10 years of experience were most accurate at 74.2%, beating every AI model. The best modern foundation model reached 72.2%, beating physicians with under 3 years of experience and matching those with 3 to 10 years, but still falling short of the experts. An older first‑generation model trailed all human readers at 56.7%. Why it matters
The story isn't that AI lost to the experts (for now). It's that the foundation model already beats junior clinicians. In settings without a specialist on hand, the value is in raising the floor, not the ceiling (yet). I'd want a tool like this deployed as a backstop for less-experienced clinicians, with experts still owning the hard cases. As models train on those hard cases, the floor will continue to be raised. Researchers at UCSF and UC Berkeley prospectively deployed Mirai, an AI breast-cancer risk model, at Zuckerberg San Francisco General, a safety-net hospital. Across more than 4,100 screening mammograms, the model scored each woman's 1‑year risk in real time and flagged 525 women (~12.7%) as high-risk. Those women could get same‑day reading of their mammogram and, when needed, same-day diagnostic imaging or biopsy. That cut the wait for diagnostic evaluation from several weeks to about an hour, and biopsies for those later diagnosed with cancer from over 2 months to under 10 days. Why it matters
Safety-net patients are the ones most likely to fall out of the system between an abnormal scan and a diagnosis. Using AI to triage who needs immediate attention, and acting on it the same day, targets exactly that gap. It's a small, single-site study, but it's a model that ideally has an outsized impact on safety-net hospitals looking to do a lot with little in serving the underserved. Read the npj Digital Medicine study → | Read the UCSF summary → |
|
1
Key Insight
The same tools that widen gaps can also close them.
Last week this brief described the recursive care law: i.e., clinical AI clusters where resources already exist, learns from those patients, gets better for those patients, while leaving under-resourced systems to inherit models optimized for someone else's population. The worry is that AI widens the very gap it could close. This week offered a counterweight. Massive Bio put its trial-matching AI inside advocacy platforms built for Black breast cancer patients, a group clinical trials have long underrepresented. And at a safety-net hospital in San Francisco, UCSF used an AI risk model to fast-track high-risk women from mammogram to same-day diagnosis, closing the weeks-long window where vulnerable patients may drop out. Both are deliberate choices to point AI at the patients who are usually underserved. Takeaway
The technology doesn't decide who it helps; the deployment does, i.e., we do. The same model that widens disparities when it simply trains where resources are most easily available can actually narrow the disparity gaps when it's aimed at them. I'm optimistic that more AI builders will start making that aim explicit, because this week shows the tools are clearly capable of it. |
|
Know someone who'd find this useful? Share |
Who AI helps is a choice, not a feature.
3 signals, 2 studies, 1 insight for health leaders
Jun 6, 2026
•
4 min read