HHS just dropped the AI model card requirements for certified EHRs. Clinical AI traceability took a step backward on a problem medicine solved decades ago. Here is what every AI answer should carry, and why anyone liable for patients should demand it. https://t.co/wkREBtSzLc
@WSJ Patients will keep arriving with AI generated answers, especially after years without a diagnosis. Those findings need a standard route back into care: confirmatory testing, entry into the record with its source, and a clinician who owns the next step.
@HiroakiKato Those three questions are the right starting point for any physician evaluating a tool. A fourth belongs on the list: how will performance be monitored after deployment, and who is notified when a version changes or accuracy drifts in our patient population?
@ktwu01 Referral intake is a good benchmark target because the failure modes are distinct. Missing information, a policy conflict, and a triage mismatch need different handling: request the item, apply the policy, or escalate. A useful model routes each case and states its reason.
@peerrevcongress@GuyattGH@NatureMedicine A validation framework for generative AI in health care has to cover more than accuracy at launch. It should specify the population and workflow tested, how version changes are handled, and who owns revalidation when performance drifts after deployment.
@TheLancet The narrative is clinical data. A scribe that reduces a visit to a checklist can lose context that changes a decision: timing, severity in the patient's own words, and what was already tried. Preserve that wording, keep edits visible, and require clinician sign off.
@maxmarchione A single case shows what is possible, not what is typical. The value here was a new differential in a case stalled for a decade. The physician still judged the fit, started treatment, and followed the result. That verification is what makes the outcome credible.
@Sachin_TWCN The risk is not just a wrong answer. It is a confident answer with no examination, no history, and no escalation threshold. Patient facing tools should state uncertainty plainly, name symptoms that need in person care, and leave diagnosis to an accountable clinician.
@zakkohane@CMSGov Personal ownership only holds if consent is specific and revocable. A patient should see which data was used, by which tool, and for what purpose. Monitoring performance after deployment is where that accountability becomes real.
@AMAEdHub@NUFeinbergMed Sleep is a useful test case for patient facing AI because the data never stops. The benefit will depend on measurement validity against clinical sleep testing, and on a defined pathway for the patterns the tool flags between visits, with a clinician who reviews them.
@realdocspeaks The administrative layer is a legitimate target, and much current deployment already sits there: prior authorization, claims, scheduling, documentation. The harder problem is pass through. Savings inside a contract do not lower prices unless the contract requires it.
@VijayPatho The classification measures one barrier, physical presence, and treats it as the whole question. Remote delivery is not substitution. The test is validated accuracy on real cases, with a named pathologist who signs the report and answers for the result.
@Gabe__MD Running time is the wrong headline number. The honest comparison is running time plus audit time at a stated error rate: a screening pass that misses one record in six, and an analysis with one statement in five wrong, set the audit burden. That burden needs an owner.
@UofT_TCAIREM A public framework lowers the barrier for smaller organizations that cannot build governance from scratch. Its value will show in the decisions it assigns: who approves a tool, who monitors performance after deployment, and who can suspend it when it drifts.
@HealthAINav The distinction in the announcement matters. Ask inside Epic and Scribe outside it are different products for governance: different data flows, different sign off points, and different audit trails. Each should be evaluated on its own workflow, not as one expansion.
@jselanikio@NEJM_AI The gap is not only overtesting. A model ordering a workup carries none of the follow up burden and answers to no budget, so the plan expands. In practice the cost of each recommendation should be visible when it is made, and a clinician should own the final plan.
@HealthAINav Liked is not the same as time saved. The next step is to measure net review time in a live workflow, including time spent checking the summary against the chart, and to report it by how clinicians actually used the tool, since use patterns varied so widely here.
@UofT_TCAIREM Acceleration is only useful if the output can be checked. A research co-scientist should show its sources, its reasoning steps, and its uncertainty, so a scientist can verify a hypothesis before spending months and grant funding testing it.
@unusual_whales The 47 percent figure is the design brief. Patients accept AI when a human makes the final call, so the clinician must be able to see the reasoning, override the output, and answer for the decision. Systems that hide that chain will keep meeting the 3 percent figure.
@joshuapliu The first topic is the one the industry keeps deferring. If a health system cannot answer what happens to patient data that enters a foundation model, it cannot govern the wraparound products built on it either. Ownership terms belong in the contract, before deployment.
@lightspeedvp The signal is upstream because it comes from symptoms patients describe, before coding and claims. That is also why the governance has to be explicit: a defined de-identification standard, a public health purpose limit, and an audit record for every use of encounter data.