|
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.
|
On June 23, Pathway Labs announced FDA clearance of EchoNext, a tool that reads a standard 12‑lead ECG to flag patients who likely have hidden structural heart disease and should get an echocardiogram. It is cleared for 6 conditions, including right and left‑sided heart failure, valve disease, and pulmonary hypertension, and was trained on more than 700,000 ECG‑echocardiogram pairs from NewYork‑Presbyterian. Across studies spanning 20+ hospitals and 500,000 patients in the US and Canada, the company says it identified structural heart disease more accurately than cardiologists, including cardiologists using AI assistance.
So what?
An ECG is one of the cheapest, most common tests in medicine, and turning it into a screen for structural heart disease could surface disease far earlier in the course, reduce disease burden and ideally also lower downstream costs.
Read the announcement → | Read the STAT coverage →
On June 25, UpDoc announced what it calls the first FDA clearance for a Software‑as‑a‑Medical‑Device built on patient‑facing large language models. Working within physician‑approved parameters, it handles tasks that used to need a clinician's hands‑on involvement, like adjusting a patient's medication, ordering follow‑up labs, and coordinating with the care team; its lead example is titrating insulin for type 2 diabetes. The system is being deployed at Cleveland Clinic, Allegheny Health Network, and UCSF Health, and raised an $18 million seed round whose investors include Eli Lilly, the American Diabetes Association, and Mayo Clinic.
So what?
The FDA clearing a patient‑facing model to take clinical actions is a meaningful line to cross. The guardrails making that possible are worth studying.
Read the announcement →
On June 22, Governor Dan McKee signed laws that restrict therapy and psychotherapy to licensed professionals, which effectively bans AI‑delivered therapy and bars AI designed to simulate emotional attachment or act as a mental‑health companion. A companion law requires any provider using AI to document a visit to tell patients and to review the AI‑generated notes for accuracy. Rhode Island joins a small but growing group of states writing AI rules specific to behavioral health and clinical documentation.
So what?
This state‑by‑state patchwork of AI regulations will be a real source of compliance confusion and cost for AI companies, health systems, and health plans that operate across state lines.
Read the summary →
|
An independent team benchmarked two commercial clinical AI tools, OpenEvidence and UpToDate's Expert AI, against three frontier general‑purpose models (GPT‑5.2, Gemini 3.1 Pro, and Claude Opus 4.6). They tested across 500 MedQA questions, 500 HealthBench items, and 100 real, de‑identified physician queries, with 12 clinicians producing 1,800 blinded ratings. The general models outperformed the specialized clinical tools on all 3 evaluations; on the real physician queries, the clinical tools did only about as well as Google's AI Overview.
Why it matters
The "clinical‑grade" tools many systems pay for are supposed to beat a general chatbot, and in this testing they didn't. The takeaway: evaluate these tools independently, on your own data with your own queries, before you buy based on an external claim.
Read the study →
A separate Nature Medicine paper put flagship general‑purpose models, including GPT‑5, Gemini 2.5 Pro, and GPT‑4o, through a series of adversarial stress tests on standard medical‑imaging question sets. The models often answered correctly even when the medical image was removed, flipped their answers after trivial changes like shuffling the options or adding an "unknown" choice, and produced fluent, confident reasoning for answers that were wrong. The gap between high benchmark scores and real robustness is wide.
Why it matters
A top score on a medical‑AI leaderboard can come from a model exploiting quirks of the test rather than actually reading, understanding and processing all of the data. Ultimately, our testing rigor needs to stay ahead of the rigor of the models themselves. This is a nice start.
Read the study →
|
A clearance and a benchmark score aren't the same as readiness
This week two AI tools cleared the FDA: EchoNext, a well‑validated ECG screener, and UpDoc, a patient‑facing model cleared to take clinical actions inside set guardrails.
Two recent Nature Medicine studies complicated the picture. The specialized "clinical‑grade" tools many systems pay for lost to general‑purpose chatbots, and every frontier model proved brittle once its medical benchmarks were nudged even slightly.
Takeaway
A clearance certifies a specific, narrow use. A benchmark score certifies performance on a fixed test. Neither one certifies that a tool holds up in the messy reality of a clinic.
|
|