ARPA-H is funding a patient-facing clinical AI — and a second AI whose entire job is to catch the first one being wrong
September 21, 2026 · 6 items
ARPA-H is funding a patient-facing clinical AI — and a second AI whose entire job is to catch the first one being wrong
Healthcare IT News · Sep 21, 2026 · September 21, 2026Practice operations
ARPA-H committed up to $33.7M in year one of a four-year, $62.7M program — ADVOCATE (Agentic AI-Enabled Cardiovascular Care Transformation). Three companies build the patient-facing agent: Atman Health (LLM clinical decision engine, voice-first interface), Tempus AI (continuous monitoring inside its Olivia patient app), UpDoc (a clinician-built rules system that validates every proposed action against approved protocols before execution).
Stanford builds something different: a disease-agnostic supervisory agent that watches the clinical agents after deployment for unsafe recommendations and out-of-distribution behavior, through a three-stage compute-escalating pipeline — outlier filtering, rule-based screening, deep-research auditing — producing "inspectable per-claim rationales."
The cohort must submit a first-of-its-kind FDA authorization package within 24 months. Dr. Rick Abramson, director of FDA's Digital Health Center of Excellence, is on record in the announcement: "Patient-facing agentic AI is one of the most consequential frontiers in digital health."
Deployment is not a pilot: Kaiser Permanente embeds the agents in EHR workflows across 21 medical centers and 260+ clinics; Duke runs multi-site validation across five health systems and rural sites on both Epic and Oracle. ARPA-H's stated stakes: 200,000+ preventable US cardiovascular deaths a year, nearly half of US counties with no cardiology provider, an estimated $28B in annual savings.
Why it matters for an independent practice: P1/P7. The regulatory template for autonomous patient-facing AI is being written right now, and its centerpiece is not the clinical agent — it is the accountability layer built to audit it. Any practice that adopts patient-facing AI in the next two years will be asked the same question this program is designed to answer: who is checking it, and can they show their work per claim? Build the audit trail before you need it, not after a board asks.
An MSC product that "worked" for frailty just missed its endpoint in infants — and a stem-cell biologist asks the question every regen practice should be ready for
The Niche / Paul Knoepfler, UC Davis · Sep 20, 2026 · September 20, 2026Regenerative medicine
Longeveron posted topline results from ELPIS II, a Phase 2b trial giving its mesenchymal stromal cell product laromestrocel to 40 infants with hypoplastic left heart syndrome as an adjunct to surgery. It failed to meet its primary endpoint — no significant benefit reported. The data did suggest the approach is safe. The stock fell hard on the miss.
Knoepfler's read is the useful part, and it is a warning about evidence hierarchy, not about this one company: he was already unconvinced by Longeveron's earlier Cell Stem Cell frailty paper, flagging "some odd change in the placebo group's performance late in the trial," and he calls the surrounding coverage — Nature included — a top contender for his 2026 "Screamers" science-hype award.
The question he keeps returning to is the one a prospective patient will eventually ask out loud: "why would one MSC product be effective for both frailty and a heart issue?" Investigational MSC therapies, he argues, are marketed on an implicit premise that MSCs treat nearly anything.
:warning: Source note: The Niche is a Candidate, added Sep 14 specifically to replace the orthobiologics coverage lost when RyOrtho gated its Biologics vertical. This is its second supplied item in eight days. This post is a "weekly reads" roundup; the Longeveron topline release inside it is the dated announcement being reported, and the rest of the roundup is not the item.
Why it matters for an independent practice: P3/P2. This is the exact failure mode that damages an independent regen practice's credibility: a promising early result gets amplified by tier-one press, patients arrive quoting it, and then the confirmatory trial misses. A practice whose marketing leans on "studies show" inherits the retraction risk. The defensible position is the narrow one — name the indication, name the trial, and say plainly where the evidence stops.
FDA added a warning to Kybella for what happens when it is injected where it was never approved — 117 unresolved nodules, persisting an average of 143 days
FDA's September 15 labeling update for Kybella (deoxycholic acid) warns of serious reactions when the drug is injected outside the approved submental region — that is, anywhere other than under the chin. Its review identified 129 adverse-event cases involving other sites and 117 unresolved nodules or masses, persisting an average of 143 days where duration was reported.
The agency is careful about what the numbers do and do not say, and so should anyone quoting them: nodules occurred at both approved and other sites, and these potentially overlapping report counts do not establish incidence. They are a signal, not a rate.
The same issue carries a second dated action worth knowing about even though it fell outside our window: FDA posted Centric Compounding's nationwide recall on Sep 13 — six lots of glutathione 200 mg/mL, Myer's Cocktail and Tri-Immune Boost — traced to elevated endotoxin in glutathione API, after nine reported reactions including fever, chills, chest pain and vomiting. Distribution reached patients' homes, not just prescribers.
:warning: Two disclosures. (1) 503Pharma is a Candidate, proposed yesterday, Sep 21, against the longest-running hole on this board — Lens B had been held fifteen consecutive days. This is its first supplied item, one day after being proposed. (2) The FDA action is seven days old — the outer edge of our window, and the freshest genuinely on-lens compounding item available. Nothing newer existed; the source publishes weekly.
Why it matters for an independent practice: P1/P2. Off-label injectable aesthetics is adjacent to a lot of what regen and aesthetic practices actually sell, and this is now a documented, labeled risk rather than a rumor. Two concrete moves: check whether any clinic in the network administers deoxycholic acid outside the submental area, and treat the consent conversation as a marketing asset — the practice that can explain the 143-day nodule finding before a patient finds it themselves is the one that keeps the patient.
Your AI-search referral number is the floor, not the count — and the fix starts with one line in your intake form
Search Engine Land · Sep 21, 2026 · September 21, 2026Patient acquisition
The core claim: GA4's AI Assistant channel finally lets you see LLM traffic, but what you see is the floor, not the norm. Three named reasons — most LLM interactions end with no click at all; LLM mobile and desktop apps strip the referrer header, so those visits land as "direct"; and people start on phone and finish on desktop, losing the referrer in between.
Her framework has three layers. (1) Direct attribution — LLM traffic and leads you can actually see in GA4 or the CRM. Her line on it is the one worth keeping: "a small number attached to closed-won and revenue beats a large number attached to nothing." (2) Influenced attribution — direct traffic and branded organic rising (people copy your domain out of the answer, or type your brand into Google days later), plus shorter sales cycles and better win rates because objections got handled during the LLM research. (3) Future moat — share of voice, citation rate and sentiment: is the model mentioning you, and is it saying the right thing?
The single cheapest instrumentation in the whole piece, and it costs one afternoon: add "LLMs" / "AI assistant" to the "How did you hear about us?" dropdown on every conversion form, and have intake staff ask it on the phone. Self-reported attribution is how you see the dark funnel at all.
She also names the honest timeline: track these monthly and treat it as an experiment — six months or more of combined data before the metrics tell a coherent story.
:warning: Conflict disclosure (standing rule for this outlet). Search Engine Land is Semrush-owned, and the per-item scan fired: this article carries an inline Semrush promo mid-body — "Be the brand AI recommends," linking to `semrush.com/ai-seo`. The framework itself is tool-agnostic and names GA4, GSC, HubSpot and Salesforce, not Semrush.
Why it matters for an independent practice: P5/P3. This is the most directly liftable item on the board for MMR. Every client asking "is the AEO work paying off?" is asking a question nobody can answer from GA4 alone, and the three-layer split gives you a defensible reporting structure instead of a shrug. The dropdown change is a same-week deliverable you can ship across the whole client roster — and six months from now it is the only first-party data any of them will have.
An accountant shut his own firm to build the thing that replaces it — 0% medical, and the pattern is coming for every service layer a practice buys
TechCrunch · Sep 21, 2026 · September 21, 2026Buildable AI
Ahad Ali ran a 20-person accounting firm handling 2,000+ returns a year — bookkeeping, sales tax, the lot. He left it. His conclusion, in his own words: "I realized that small businesses don't need better accounting software. They need less accounting software. They need something that just does it for them."
His product Tabby pulls live bank data through Plaid and uses AI to maintain a real-time P&L dashboard, with the visible software layer deliberately shrinking toward zero: "You can say QuickBooks digitized bookkeeping, but we want to make [that visible software layer] nonexistent."
The traction numbers are small and real, which is the useful part: fourteen months in, 5,500 small businesses, roughly $100K ARR, a team of seven, raising a $1M pre-seed. He is going direct to the end user wherever possible rather than selling through firms.
The strategic bet: turn bookkeeping into a flat monthly utility charge — "on the order of internet access or payroll services." He is up against QuickBooks, the Sequoia-backed Rillet, and the frontier labs' own finance tools, and is betting the market is big enough anyway: "There's 30 million more businesses."
Why it matters for an independent practice: P4/P5. Nothing here is medical, which is the point. Watch the shape: an insider with 2,000 reps of the manual work builds the thing that eliminates it, and the category becomes a utility line item. That shape is already pointed at medical billing, prior auth, intake and — squarely — at agency marketing services. Two readings for MMR. Defensive: the deliverables a client can eventually get as a $99/month utility are the ones to stop selling by the hour. Offensive: Atiba has the 2,000-reps insider knowledge in his category, which is exactly the asset Ali monetized — the productizable part of MMR is the judgment, not the labor.
Nate Herk dropped a 5-hour Codex course for people who don't code — and the last chapter is how to price and sell what you build
Nate Herk | AI Automation · uploaded Sep 21, 2026 · September 21, 2026Buildable AI
Date-verified from the watch page's `uploadDate` meta — 2026-09-21T05:31:48-07:00 — not a channel-index relative label. ~24K views at time of check. Note the title is being A/B tested: it renders as both "Codex for Non-Coders (5+ Hour Course)" and "Build & Sell with Codex (5+ Hour Course)" on the same video ID.
What it actually covers, from the video's own description: building "your AI operating system and second brain" with GPT-6 Astra, creating reusable skills, generating websites and branded deliverables, and generating and editing video — all without a coding background.
Then the operational half: browser use, voice mode, hosting your automations, and Codex vs. Claude Code — a direct comparison rather than a single-tool pitch.
It closes on how to price and sell AI solutions to clients, which is the part that separates this from a tooling walkthrough. Herk runs the AI Automation Society community; the build-and-sell framing is the channel's whole thesis.
Why it matters for an independent practice: P5/P4. This is the standing preference in its purest form — a no-code builder walkthrough that never mentions medicine. The medical read: "branded deliverables and websites, generated" is the MMR production line, and "reusable skills" is the difference between doing a client deliverable once and doing it fifty times. The hosting and browser-use chapters are what turn a demo into something that runs at 6am without you. And the pricing chapter is the one to watch with a competitor's eye — it is teaching a million subscribers to sell the same service layer.
