Weekly review: three roads to the bedside were walked at once — one skips authorization, one declines to be a device, and one took 46 days from clearance to first incision
Taken a day at a time, this week looks like unrelated headlines. Stack the seven days and the shape appears: the same class of generative-AI products reached patients through three entirely different legal identities in the same week. FDA's TEMPO pilot put four products in front of Medicare patients before any marketing authorization; OpenAI wired ChatGPT into Epic charts on September 1 by simply not claiming to be a device at all; and Stryker's SportSuite Vision walked the full De Novo route, cleared on July 17 and reaching its first case at Duke Health on September 1. The evidence bar across those three roads differs by an order of magnitude — and the patient is never told which road they are standing on.
01 — Top Stories
① TEMPO expands to four products: generative-AI devices reach Medicare patients before any marketing authorization
FDA's TEMPO (Technology-Enabled Meaningful Patient Outcomes) pilot expanded this week, per STAT News (Sep 3). The FDA's own participant list now names four: Limbic's "Unpacked" (voice-AI delivery of CBT for depression and anxiety in Medicare beneficiaries), SonderMind's SACA (adults 22 and older), Cadence Solutions' HypertensionOS (clinician-supervised antihypertensive management in Stage 2 hypertension), and the Dexcom Glucose Health Program (metabolic and nutritional monitoring, prediabetes screening). The names matter less than the pilot's own terms: participants may request FDA enforcement discretion over "premarket authorization requirements, investigational device exemption (IDE) requirements, and requirements under 21 CFR parts 50 and 56." Part 50 is informed consent. Part 56 is IRB review.
For twenty years the only lawful route for an unauthorized device to touch a patient was a clinical trial — and the price of a trial is informed consent and IRB oversight. TEMPO folds both into the discretionary zone, and buys "real-world evidence" with them. This is not a question of review speed; it moves the burden of proof from premarket to postmarket and asks the enrolled patients to carry the gap in between. FDA says it expects up to roughly ten manufacturers in each of four clinical areas — about forty in total — so this is not a one-off exception but a deliberately built parallel track.
The STAT piece sits behind a paywall; the four-name list here comes from FDA's own page. TEMPO's enforcement discretion is something participants may request, not something automatically granted, and FDA does not publish case-by-case which requirements each firm was relieved of. The 34-million figure for Medicare beneficiaries under the ACCESS model comes from Fierce Healthcare and describes the model's potential population, not actual enrollment.
② ChatGPT can now read Epic charts — and at no point does it claim to be a device
OpenAI announced on September 1 that ChatGPT for Healthcare customers can connect Epic environments to ChatGPT, letting clinicians pull visit notes, lab results, medications and specialist documentation inside the chat interface. A Healthcare Public Data plugin shipped alongside it, wired to nine official sources including ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed and PubMed. OpenAI says physicians from 60 countries and 26 specialties rated more than 700,000 model responses, that "physicians rated 99.1% of responses safe across all use cases," and that a separate evaluation found "more than 93% of responses were rated as having 'good' or better accuracy." Both TechCrunch and healthsystemCIO describe the access as read-only.
This only lands when read next to ① . The TEMPO road at least has FDA standing beside it, collecting outcomes and retaining the power to pull a product. The ChatGPT road never crosses a regulator's desk. Read-only is the load-bearing design choice: no orders, no diagnostic claim, therefore no device definition, therefore no premarket evaluation of any kind. The result is three layers of AI stacked on the same chart — Epic's own agents, the incumbent ambient scribe, and now ChatGPT — with the question of who owns which output left to hospital CIOs to sort out.
Both the 99.1% and 93% figures come from OpenAI's own announcement. They are physician ratings of responses — not clinical outcomes — and are not independently audited. The operational definition of "safe," the raters' independence, and the sampling method are not disclosed.
③ The first FDA-authorized surgical app for Apple Vision Pro: De Novo on July 17, first case on September 1
Stryker's SportSuite Vision received De Novo authorization on July 17, 2026, and reached its first case on September 1 at Duke Health, where orthopedic surgeon Chad Mather III performed a hip arthroscopy for femoroacetabular impingement with labral repair. The system is a video see-through augmented reality head-mounted display that layers arthroscopic video, HipCheck, HipMap and CT imaging into the surgeon's field of view while the rest of the OR staff stays on conventional monitors. Mather said it "helped create a more comfortable, streamlined OR setup while keeping the information I needed in view." STAT News called it a "surgical cockpit."
On its own this is a modest story, but it is the week's best control group. De Novo exists for devices with no predicate to point at; Stryker walked it end to end, and 46 days separated authorization from the first incision — the time it takes to validate, train, schedule and roll out inside a hospital. The four TEMPO products and the ChatGPT–Epic connector skipped those 46 days entirely, because they are not on this road. The speed difference is the lesser point. The real one: this is the only road where anyone can go back afterwards and find out what FDA actually reviewed.
④ The week's hardest evidence was not American: a 10,333-patient prospective trial in which AI surfaced 51 liver lesions humans had missed
The LiON study, published in Nature Medicine on August 19, trained on 6,443 patients, validated retrospectively across 22,251 patients in a multicenter cohort, then ran a prospective single-arm trial in 10,333 patients in routine practice across eight Chinese hospitals. Retrospective validation AUC was 0.975 (95% CI 0.971–0.979), falling to 0.952 (0.942–0.961) in the real-world cohort — 0.971 in patients with hepatic steatosis, 0.924 in cirrhosis. The prospective trial met its primary endpoint, and AI–human collaboration identified 51 previously overlooked lesions, 15 of them malignant, prompting 37 amended radiology reports and 22 multidisciplinary team escalations.
Put that beside the week's other news: an analysis widely covered in late August found that of 1,357 FDA-cleared AI devices, only three had been tested against patient outcomes (also reported by Healio). The most solid clinical evidence of the week therefore came from a research pathway entirely outside FDA's jurisdiction. And the decay itself — 0.975 retrospective, 0.952 real-world, 0.924 in cirrhosis — is precisely the question every product that reports only internal validation should be asked.
A single-arm design has no control group. The 51 missed lesions are the output of AI–human collaboration and cannot be attributed to the model alone, nor does the design answer how many false-positive workups were generated alongside them. All eight sites are in China; the paper makes no claim that the population and imaging-equipment mix generalize to Taiwan, Europe or the US.
⑤ Once again this week, none of the money went to the model layer — it bought prior auth, distribution and risk stratification
Per Health IT Answers' September 3 financial roundup: R1 RCM agreed to acquire Humata Health (AI prior-authorization automation; terms undisclosed, expected to close at the end of Q3 2026); Cityblock Health closed a $116M Series E and acquired Homeward Health, taking the combined entity to nearly 250,000 members; WellStack acquired predictive-analytics and risk-stratification firm DeLorean AI; Savista acquired ABW Medical; and Switchboard Health raised $5M after acquiring Livara Health. R1 CEO Joe Flanagan said the deal "advanc[es] our strategy to have the most intelligent and integrated pre-bill architecture."
This is the second straight week with the same shape: not one acquisition bought model capability. What buyers bought were the narrow points money passes through — prior authorization, pre-bill architecture, rural distribution, member counts. Humata is worth buying because KFF found insurers denied 12–18% of prior-authorization requests in 2025 across Medicare Advantage, Medicaid managed care and ACA marketplace plans; that friction is the market. Who built the underlying model is, by comparison, a swappable component.
Humata's performance figures — 96% first-pass approval, 30% fewer write-offs, 83% fewer rescheduled appointments — all come from executives on both sides of the deal, with no third-party audit and no stated baseline or sample. The R1 terms are undisclosed and Cityblock's only partly disclosed.
⑥ The one window still open: FDA's generative-AI device discussion paper closes for comment on October 19
On August 18 the FDA opened a request for public comment (full discussion paper here) proposing a two-axis risk framework, competency-based premarket evaluation combining device benchmarking with clinical confirmation, risk-proportionate postmarket monitoring, and specific treatment of foundation models and agentic AI. Docket FDA-2026-N-7874; comments close October 19, 2026. CDRH director Michelle Tarver: "By inviting input from the public, we are launching a transparent process to inform development of an approach that safeguards patients."
This paper and TEMPO are two moves by the same agency in the same month, pointing opposite ways: one asks how generative-AI devices should be reviewed, while the other has already put four of them into use before review. Practically, comments filed before October 19 are likely to shape the evaluation framework Taiwanese vendors will face when exporting to the US for the next three to five years. It is the week's only node where an outsider still has leverage.
⑦ Background thread: Epic's roadmap folds all of the above into a single interface
The timeline Epic set at UGM 2026 in August started to bite this week: Ergo, an AI-native clinical interface, ships in November; Agent Factory offers 120-plus pre-built agents with no coding and broadens in 2027; and Cosmos Curiosity, a generative forecasting model, is slated for March 2027. The scale numbers: Cosmos spans 320 million patients and 23 billion clinical encounters, Care Everywhere exchanged 914 million records in June alone, and MyChart Central is live in all 50 states across 346 organizations serving 115 million patients. Berlin's Charité became Epic's first German customer on a €200 million agreement with full implementation targeted for end-2029; Epic now operates in 17 countries. Fierce Healthcare covered the same roadmap.
TEMPO products entering hospitals, ChatGPT reading charts, ambient scribes writing back — all three must pass through Epic's interface and data permissions. With Epic shipping its own AI-native interface in November and 120-plus pre-built agents next year, it is simultaneously everyone's channel and everyone's competitor. That is where the week's three roads to the bedside converge, and the structural reason buying distribution beats buying models.
02 — Product Analysis
Limbic "Unpacked"
Voice-AI delivery of cognitive behavioral therapy · Limbic Inc. (US/UK)
Function and position. Twenty-minute CBT sessions delivered over scheduled phone calls, with a clinician supervising each patient's care, reviewing session content and receiving real-time safety alerts. TEMPO lets it enter CMS's ACCESS chronic care model before any marketing authorization — the first AI-led mental health solution among the four participants.
- Strength : accessibility is a real differentiator. Voice delivery means it works on a landline or a flip phone — and the Medicare population is precisely the cohort app-based mental health platforms reach worst. The company says it has served 850,000 patients across a dozen US states and the UK.
- Strength : the bar to selection was not trivial. CEO Ross Harper says the application ran past 100 pages and included nine peer-reviewed studies — a thicker evidence base than most "AI therapy" products publish.
- Concern : TEMPO's enforcement discretion reaches 21 CFR parts 50 and 56 — informed consent and IRB review. Treating Medicare patients with depression and anxiety using an unauthorized autonomous therapeutic, in a prospective deployment, without the standard human-subjects protections, is a trade worth reading FDA's own pilot description word by word before judging.
- Concern : the regulatory wind blows the other way. Several states are moving to restrict autonomous AI therapy, and Limbic's relief is federal while the practice of psychotherapy is regulated mostly at state level. The 850,000-patient figure is company-reported and not independently verified.
ChatGPT for Healthcare — Epic connector
Enterprise chart retrieval and public-data lookup · OpenAI (US)
Function and position. Lets ChatGPT for Healthcare customers, and ChatGPT Enterprise customers with a BAA, connect Epic environments into the chat interface for read-only retrieval of visit notes, labs, medications and specialist documentation. US individual ChatGPT for Clinicians users get the public-data plugin only, with EHR integration excluded.
- Strength : putting the chart and nine official public sources (ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, PubMed and others) behind one query interface removes the cost of switching between systems — the most direct pain point incumbent medical retrieval tools address.
- Strength : the evaluation scale is unusual for this category — physicians across 60 countries and 26 specialties rated more than 700,000 responses. Even discounting the conclusions, that sample exceeds what most medical AI vendors have ever published.
- Concern : no device claim means no premarket evaluation. Clinicians will use this to read real patients' labs and medications and form judgments, with no regulator reviewing it beforehand — every piece of safety evidence it has is the piece the vendor chose to publish.
- Concern : layered accountability is unresolved. Epic's own agents and an ambient scribe already sit on that chart; adding ChatGPT means that when something goes wrong, there is no standard way to trace which layer produced it. The flip side of "99.1% safe" is nine unsafe responses per thousand — not a remainder you can round away inside a patient chart.
03 — Companies & Competition
| Company | Recent state & numbers | Position & moat |
|---|---|---|
| Epic EHR incumbent; where every road converges |
Cosmos spans 320M patients and 23B encounters; Care Everywhere exchanged 914M records in June alone; MyChart Central covers 50 states, 346 organizations, 115M patients; Berlin's Charité signed a €200M deal as Epic's first German customer (UGM 2026 recap). | The moat is data permissions and workflow position, not models. The weakness: it is channel and competitor at once — once Ergo ships in November, partners start calculating whether they are the next feature to be absorbed. |
| OpenAI Entering the clinic by not being a device |
Shipped the read-only Epic connector and a nine-source public-data plugin on Sep 1; self-reported physician ratings of 700K+ responses across 60 countries and 26 specialties, 99.1% rated safe (company announcement, not independently audited). | The moat is general model capability plus enterprise contracts and BAA coverage. The weakness: its clinical position rests on not claiming — and if FDA's generative-AI framework lands and redraws the clinical-decision-support boundary, that position can expire overnight. |
| OpenEvidence Physician-facing retrieval; shipped its own model family this week |
Launched a new family of clinical models on Sep 3 (STAT News, paywalled), following a $250M Series D. | The moat is physician habit and journal licensing. The weakness is blunt: ChatGPT now reads the chart too, and OpenEvidence has no Epic-level data permission to fall back on — only depth and citation quality. |
| Limbic First AI-led mental health entrant in TEMPO |
Selected for TEMPO, entering the ACCESS model pre-authorization; self-reports 850,000 patients across a dozen states and the UK, with nine peer-reviewed studies in its application (Fierce Healthcare). | The moat is a first-mover regulatory position plus voice accessibility. The weakness is that the position is borrowed — TEMPO is a pilot, and state-level restrictions on autonomous AI therapy are tightening. |
| R1 RCM Revenue cycle; buying the prior-auth chokepoint |
Agreed to acquire Humata Health, terms undisclosed, closing expected end of Q3 2026; Humata self-reports 96% first-pass approval, 30% fewer write-offs, 83% fewer rescheduled appointments (Fierce Healthcare, unaudited). | The moat is integration depth in the pre-bill stack plus existing hospital contracts. The weakness: the value depends on payer friction persisting — simplify the coverage rules and this market shrinks on its own. |
| Stryker The one that walked the full De Novo |
SportSuite Vision received De Novo authorization July 17 and completed its first hip arthroscopy at Duke Health on September 1 (Stryker press release). | The moat is hardware, the surgical instrument portfolio and surgeon relationships — and the De Novo itself now sets the bar for followers. The weakness: headset adoption in the OR moves far slower than software, and it is bound to a single consumer hardware platform. |
Close the section in one sentence: the contest has left the "whose model is better" layer. This week's winners were competing over who can touch a patient without review, and who sits at the chokepoints where money and data pass. Model capability is the entry ticket; legal identity and channel position decide the match.
04 — Taiwan Angle
(1) Taiwan's default is disclose-first, consent-first, human-confirms-last; what the US demonstrated this week is deploy-first, gather-evidence-later. The MOHW's Guidelines on the Use of Generative AI in Healthcare Institutions, issued May 29, 2026, define generative AI explicitly as an assistive tool, require clinicians to make the final confirmation and review, mandate continuous monitoring and incident reporting, and require that patients be told the scope of AI involvement and give informed consent (Lee and Li's clause-by-clause summary). The guidelines name six risk categories including output hallucination, prompt-injection attack, and erosion of clinical judgment through over-reliance. Set that against TEMPO's enforcement discretion over 21 CFR parts 50 (consent) and 56 (IRB): the two defaults point in opposite directions.
(2) Where a product lands changes procurement directly: a "not a device" product like the ChatGPT–Epic connector falls outside TFDA device regulation in Taiwan and lands instead under institutional governance guidance. The guidelines' three-stage framework requires a designated unit to identify and grade risk before adoption, to assess integration with EMR/HIS/PACS, and to phase in high-risk applications. In practice: a US hospital buying such a tool meets resistance mainly from its CIO and compliance office, while a Taiwanese hospital carries an explicit additional duty to inform the patient. For vendors selling into both markets that is a real cost — and the local projection of this week's three divergent roads.
(3) One concrete action for Taiwanese vendors: FDA's comment window is open until October 19. The two-axis risk framework and competency-based premarket evaluation proposed in the discussion paper will, if finalized, become the evaluation vocabulary Taiwanese AI devices must speak when exporting to the US. The MOHW already operates three smart-healthcare centers and, per United Daily News, is running a "333 policy" with matching national budget. Packaging the real-world validation data already accumulated domestically into a formal comment is cheap — and what it shapes is the export bar for the next three to five years.
05 — Further Reading
-
TEMPO for Digital Health Devices Pilot — FDA (Digital Health Center of Excellence)
Put down every secondhand account of TEMPO and read this page first. It states plainly which requirements the enforcement discretion covers, how many firms are expected in each of the four clinical areas, and what qualifies a product — details news coverage almost universally compresses away.
-
Considerations for the Regulation of Generative AI-Enabled Medical Devices — FDA discussion paper (2026-08-18)
The actual document you can still comment on before October 19. Read the section on foundation models and agentic AI in particular — it is the fullest public thinking any regulator has published on models that call tools on their own.
-
Large-scale AI-guided liver malignancy diagnosis: multicenter study and a single-arm trial — Nature Medicine (2026-08-19)
Read the methods, not the conclusion. The decay curve — 0.975 retrospective, 0.952 real-world, 0.924 in cirrhosis — is the reference shape for evaluating any imaging AI, and the best ruler to hold up against a vendor that reports only internal validation.
-
How to meaningfully evaluate AI in clinical medicine — Nature Medicine
After a figure like "only three of 1,357 cleared AI devices were tested against patient outcomes," this is the piece that lays out what a meaningful evaluation standard would look like. Worth reading before your institution's AI committee next meets.
-
ChatGPT's Epic Integration Stacks a Third AI Layer on the Chart, Leaving CIOs to Sort the Overlap — healthsystemCIO (2026-09-01)
A rare piece written from the buyer's chair. It handles the question every vendor announcement this week stepped around: with three layers of AI on one chart, how do you trace and apportion responsibility when something goes wrong.
06 — References
- Participants Selected for TEMPO for Digital Health Devices Pilot. U.S. Food and Drug Administration, 2026. fda.gov
- TEMPO for Digital Health Devices Pilot. U.S. Food and Drug Administration, 2026. fda.gov
- FDA pilot offers generative AI medical devices a path to patients before they are authorized. STAT News, 2026-09-03. statnews.com
- Limbic selected to FDA pilot, will test voice AI therapy with Medicare enrollees. Fierce Healthcare, 2026. fiercehealthcare.com
- FDA's TEMPO Pilot Brings In Smaller Digital Health Firms With Three New Participants. Medtech Insight (Citeline), 2026. insights.citeline.com
- Healthcare organizations can now connect EHR and additional industry data to ChatGPT. OpenAI, 2026-09-01. openai.com
- ChatGPT Health adds Epic integration for clinicians to import patient data. TechCrunch, 2026-09-01. techcrunch.com
- ChatGPT's Epic Integration Stacks a Third AI Layer on the Chart. healthsystemCIO, 2026-09-01. healthsystemcio.com
- Stryker introduces first of its kind FDA-authorized surgical application for Apple Vision Pro with hip arthroscopy case. Stryker, 2026-09-01. stryker.com
- FDA clears Stryker's 'surgical cockpit' that uses Apple's VR headset, Vision Pro. STAT News, 2026-09-01. statnews.com
- Large-scale AI-guided liver malignancy diagnosis: multicenter study and a single-arm trial. Nature Medicine, 2026-08-19. nature.com
- How to meaningfully evaluate AI in clinical medicine. Nature Medicine, 2026. nature.com
- Only three of 1,357 FDA-cleared AI devices tested patient outcomes. News-Medical, 2026-08-20. news-medical.net
- Most AI tools cleared by FDA were not tested on clinical outcomes. Healio, 2026-08-21. healio.com
- Health IT Business News, Financial Edition — September 3, 2026. Health IT Answers, 2026-09-03. healthitanswers.net
- R1 acquiring Humata Health to bolster AI-powered prior authorizations. Fierce Healthcare, 2026. fiercehealthcare.com
- R1 to acquire Humata Health for AI prior authorization push. Becker's Hospital Review, 2026. beckershospitalreview.com
- Health Tech Funding Surges Across August. DistilINFO, 2026-09-02. distilinfo.com
- FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices. U.S. Food and Drug Administration, 2026-08-18. fda.gov
- Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback. U.S. Food and Drug Administration, 2026-08-18. fda.gov
- Holland & Knight Health Dose: September 1, 2026. Holland & Knight, 2026-09-01. hklaw.com
- Epic UGM 2026 Recap: Cosmos Curiosity, EpicOps and an Expanding AI Footprint. HealthTech HotSpot, 2026. healthtechhotspot.com
- Epic expands AI ambitions with agent platform, Cosmos-powered predictions and workflow automation. Fierce Healthcare, 2026. fiercehealthcare.com
- OpenEvidence launches new family of AI models for clinicians. STAT News, 2026-09-03. statnews.com
- OpenEvidence clinches $250M Series D, rapidly growing its reach with doctors. Fierce Healthcare, 2026. fiercehealthcare.com
- 衛福部首發醫療 GenAI 指引 六大風險、九項要點建立治理框架. 環球生技月刊, 2026. news.gbimonthly.com
- 衛福部頒布「醫療機構應用生成式人工智慧指引」. 理律法律事務所, 2026. leeandli.com
- 臺灣智慧醫療三大中心. 衛生福利部. aicenter.mohw.gov.tw
- 高醫大論壇揭示 AI 醫療新局！衛福部推「333 政策」. 聯合新聞網, 2026. udn.com
- AI Model Releases: September 2026 Tracker and Dated Ledger. Digital Applied, 2026. digitalapplied.com