First map of the regulatory vacuum: Brussels pushed the high-risk deadline to 2028, and the FDA answered with 30 pages and 26 questions on how to examine a generative device
Two things happened this week that look opposite and are the same thing. August 2 was supposed to be the day the EU AI Act's high-risk obligations bit. Instead, Regulation (EU) 2026/1744, the Digital Omnibus on AI — in the Official Journal on July 24, in force July 27 — moved them: December 2, 2027 for stand-alone Annex III systems, August 2, 2028 for AI embedded in regulated products such as medical devices. In the same window the FDA published a discussion paper on generative AI-enabled devices (docket FDA-2026-N-7874, comments due October 19) that states plainly it is not draft guidance, proposes no policy change, and does not address legal authority. Binding deadlines slide two years out; non-binding technical detail runs to thirty pages. The centre of gravity in AI regulation has moved from when you must comply to what the exam actually looks like.
01 — Top Stories
The FDA sets the first exam paper for generative AI devices: 30 pages, 26 questions, comments close October 19
On August 18 the FDA's Center for Devices and Radiological Health and its Digital Health Center of Excellence jointly issued a discussion paper on regulatory considerations for generative AI-enabled medical devices, opening docket FDA-2026-N-7874 for comment until October 19, 2026. The paper runs to roughly 30 pages and poses 26 discussion questions across risk assessment, premarket evaluation, postmarket monitoring, foundation models and agentic systems. Its lineage traces to the November 2024 Digital Health Advisory Committee meeting.
Through March this year the FDA had authorised more than 1,400 AI devices since 1995, including a record 331 in 2025 alone — but almost all of them are frozen-weight, single-task discriminative models. For the generative kind there has been no route at all. Industry says so bluntly: Ashkon Rasooli of EnGenius Solutions told MD+DI there are "currently no pathways to get a generative AI-enabled medical device on the market because FDA hasn't come to terms with quality assurance." This paper is the first attempt to fill that blank.
The FDA states in the document itself that this is not draft guidance, proposes no policy change, and does not address legal authority. Nothing you can file against was created this week. Acting Commissioner Kyle Diamantas supplies the political frame: the move aligns with an administration priority to "accelerate the delivery of AI-powered medical products to market." Read the paper as science and acceleration at once.
A two-axis grid and ten benchmarks: the FDA wants to credential models the way it credentials physicians — prompt injection included
The paper's core is two axes: activity — from non-directive information such as a risk score, through action-directing guidance such as "go to the ED", to supervised action and finally fully autonomous decision-making — and consequence severity, from limited to severe harm. Risk rises from lower-left to upper-right, and the FDA explicitly notes that adding "talk to your doctor" does not reduce directiveness if the output still pushes toward action. Premarket evaluation then splits into three stages: device benchmarking across ten elements, clinical confirmation at five ascending levels of rigour, and postmarket monitoring by periodic re-benchmarking, sampled clinician review and drift detection.
This is the first time a regulator has written LLM safety-engineering vocabulary into a premarket test list. The ten elements fall into four groups: safety (S.1 escalation speed and over-reassurance, S.2 boundary adherence and prompt injection, S.3 uncertainty calibration); clinical proficiency (E.1 guideline adherence, E.2 information gathering, E.3 quantitative analysis, E.4 communication); generalisability (R.1 robustness to paraphrase and ordering, R.2 subgroup performance across demographics and dialects); and agentic conduct (A.1 multi-step planning within safety bounds, tool-error recognition, human checkpoints before irreversible actions). Multi-turn drift out of scope counts as an S.2 failure. Two further signals matter: clinical confirmation "might not require a prospective clinical study in every case", and where the expert adjudicator is itself an LLM, the independence requirement still applies and must be validated against a human-adjudicated subsample.
The detail in this item comes from Innolitics' third-party reading of the paper, not verbatim from the FDA PDF; cite the docket original for any formal use. And the trade it floats — accepting greater premarket uncertainty in exchange for heavier postmarket monitoring — turns in practice on who funds that monitoring, a question the paper does not answer.
Six days after the paper, the FDA's digital health chief says out loud: the goal is formal guidance
On August 24, Digital Health Center of Excellence director Rick Abramson, M.D., told STAT that "Our goal is formal policy guidance. The ecosystem is expecting clarity." The route he describes is one broad guidance covering generative AI plus a set of specialty guidances on topics of particular interest or complexity. No timeline was given in the interview.
A discussion paper is legally nothing. What determines whether companies can plan is whether the road leads to guidance. Abramson's line reclassifies the paper from an academic exercise to the opening of a formal rulemaking path — an entirely different signal for a sponsor deciding today whether to fund a prospective trial. The same director also authored an August 20 white paper on digitally derived measures. Read together, one centre spent a single week on both halves of the problem: how you examine the model, and how you measure the endpoint.
The STAT piece sits behind a paywall; this item is written only from the publicly visible headline and standfirst, and the "competency-based approach" details it mentions could not be confirmed from the open portion. Note too that the absence of a timeline is itself information: in US device policy, the road from discussion paper to final guidance routinely takes years.
Brussels moved the medical-device high-risk deadline from this month to 2028 — but Article 50 transparency was not deferred and went live on August 2
Regulation (EU) 2026/1744 was published in the Official Journal on July 24 and entered into force on July 27. Stand-alone Annex III high-risk systems now fall due on December 2, 2027; high-risk systems embedded in regulated products — medical devices, machinery, aviation — on August 2, 2028. The same instrument narrows the "safety component" definition so that AI which merely improves performance or convenience, and whose failure does not endanger health or safety, no longer falls into high-risk automatically. Politically, Council and Parliament announced their simplification deal on May 7 and Member State representatives confirmed it on May 13.
The point most easily missed is the one that binds today: Article 50 transparency obligations were not deferred and have applied since August 2, 2026. Any medical chatbot interacting with patients or the public in the EU, any AI-generated patient-education content, any synthetic image, carries a disclosure duty now. What slid to 2028 is the heavy high-risk apparatus — risk management system, data governance, technical documentation, human oversight, incident reporting. For device makers there is a second layer: MDR/IVDR obligations were never loosened, and MDCG 2025-6 already sets out how the AI Act interlocks with them.
Deferred is not cancelled, and the extension is conditional on the continued development of technical standards and guidelines. Anyone treating 2028 as two years off the hook will find, as the standards land one by one, that the runway was shorter than it looked.
CMS wants a dedicated payment lane for diagnostic AI software: "Software as a Medical Service" and an O1 status indicator
In the CY2027 Hospital Outpatient Prospective Payment System and ASC proposed rule issued July 2, CMS floated a new "Software as a Medical Service" (SaMS) category flagged by a new O1 status indicator, designating 36 HCPCS codes, of which 21 would move into New Technology APCs; a further ten algorithmic laboratory analysis codes would shift from the Clinical Laboratory Fee Schedule into the outpatient system — which could expose patients to new copayments. The scope covers AI analysis of retinal images, echocardiogram-based heart failure detection, coronary blood flow estimation and similar algorithmic reads.
In the US, what decides whether an AI device has commercial life is not the 510(k) — it is whether a code exists, whether it is paid, and at what rate. A decade that produced over 1,400 authorisations with little scaled adoption owes much of that gap to reimbursement. If SaMS survives to a final rule, it concedes that an algorithmic read is a separately priceable service rather than an appendage to some existing procedure. Structurally, that may matter more than any guidance document.
This is still a proposed rule. CMS itself frames it as interim policy pending a long-term valuation methodology once claims data accumulate. Published coverage gives no payment rates or dollar figures, and none are asserted here. The OPPS final rule customarily lands around November, and may diverge from the proposal.
Ten states legislated on health AI; the federal government is preparing to sue them — the most contradictory regulatory scene in America this year
On the state side: at least seven states enacted laws in 2026 restricting insurer AI — Iowa HF 2635 (effective July 1, 2026; AI may flag prior authorisation requests but cannot be the sole basis for denial), Washington SB 5395 (June 11, 2026; only licensed professionals may deny), Alabama SB 63 (Oct 1, 2026), Colorado HB 1139 (Jan 1, 2027, which also bars AI-delivered psychotherapy), Georgia SB 444 (Jan 1, 2027), Utah SB 319 (Jan 1, 2027) and Illinois SB 3114 (Jan 1, 2028, banning automated downcoding). Holland & Knight counts ten states legislating this year across three buckets: insurer oversight, clinical-decision restrictions (Maine HB 2082, Arizona) and AI companions and chatbots (Idaho, Nebraska, Oregon SB 1546, Tennessee SB 1580, Delaware HB 191).
The federal side is pulling the other way. The executive order signed December 11, 2025, "Ensuring a National Policy Framework for Artificial Intelligence," stood up an AI Litigation Task Force at the DOJ from January 10, 2026 to challenge state AI laws on Dormant Commerce Clause and preemption theories, required Commerce to publish a list of "overly burdensome" state laws by March 11, 2026, and conditioned $42 billion in allocated broadband funding on states repealing certain AI rules. Net effect: an insurer using AI in prior authorisation faces seven mutually inconsistent state regimes this year, plus a federal government arguing those regimes are void.
The two tallies carry different dates (Becker's July 28, Holland & Knight May 26) and the count moves as sessions close, so nothing here should be read as a figure current to this week. No ruling yet exists to cite on the preemption litigation, and the executive order's practical effect remains undetermined.
TEMPO: an FDA–CMS pilot that staples clearance and reimbursement onto the same application
On August 21 the FDA refreshed its TEMPO (Technology-Enabled Meaningful Patient Outcomes) for Digital Health Devices pilot page with selection progress. Dexcom was announced as the first participant on July 22. The pilot runs jointly with CMS and plans to select roughly ten participants per clinical use area across four ACCESS model categories; those selected gain a defined payment pathway through the ACCESS model and support in generating real-world evidence.
If reimbursement is the real bottleneck, TEMPO is the same government's other answer to it: rather than waiting for CMS to open a code after the fact, it binds a payment pathway to the device during regulatory review. CDRH director Michelle Tarver frames it as "access to the right technology at the right time can be life-changing." For chronic-disease digital therapeutics and continuous monitoring — products whose evidence accrues over time — that design fits the commercial reality far better than a one-shot clearance.
As the FDA's own page stands, Dexcom remains the only publicly named participant; the August 21 refresh added no new names. "Roughly ten per clinical use area" is a plan, not an accomplished fact. And Dexcom's Glucose Health Program sits in glucose and metabolic health, a very different regulatory situation from imaging-read AI devices — do not generalise across.
Who fills the gap: America reaches for certification, Britain for a sandbox — two ways of legislating without legislation
America: the Joint Commission launched a voluntary Responsible Use of AI in Healthcare (RUAIH) certification on June 2, built on five areas — governance, data management, risk and bias reduction, monitoring and validation of safety performance, and transparency and training. It is open to any healthcare organisation, does not require Joint Commission accreditation, and expressly does not validate individual AI products. The Coalition for Health AI's governance playbooks are positioned as the framework for achieving it. Britain: the MHRA's AI Airlock regulatory sandbox won £3.6 million from the Department of Health and Social Care on April 8, spread over three years (2026–2029, £1.2m annually), with phase three extending to large language models, voice technologies and specialist diagnostics in cancer and rare disease.
When binding rules slip two years, the market does not stand still and wait. It finds substitute signals: a voluntary certificate, a sandbox slot, a vendor questionnaire. None carries legal force, yet all are hardening into de facto procurement thresholds — which is to say regulatory authority is being partly outsourced to non-governmental bodies. CHAI chief executive Brian Anderson says the certification and the playbooks are "tightly aligned on the need for responsible and transparent AI in healthcare"; MHRA executive director James Pound calls the multi-year funding "a pivotal moment for AI Airlock and safe AI advancement in healthcare." Both sentences describe the same shift.
Published coverage gives no fee, no application-window date and no count of certified organisations for RUAIH, so its real uptake cannot be assessed. AI Airlock's phase-two results were slated for summer 2026 but no published findings were available to cite as this edition closed. Both sit at the stage of mechanism-built, effect-unproven, and neither is evidence that the regulatory gap has actually been filled.
02 — Product Analysis
Dexcom Glucose Health Program
First TEMPO pilot selectee · Dexcom, Inc. (US)
Function and position. Built on continuous glucose monitoring, it uses AI-powered insights and real-time data to help users track metabolic health, receive personalised guidance, and screen for prediabetes and type 2 diabetes. The FDA named it the first TEMPO participant, and the pitch is not only the product but the fact that it simultaneously secured a route into CMS's ACCESS model payment pathway.
- Strength : clearance and payment obtained together. TEMPO explicitly promises selectees an ACCESS model payment pathway plus real-world evidence support — precisely the step at which most of the 1,400-plus authorised AI devices have stalled.
- Strength : chronic-disease data density suits real-world evidence natively. The daily volume a continuous monitor generates maps directly onto what the FDA's August 20 white paper on digitally derived measures asks for: understanding "how that patient is functioning across many moments — and whether a meaningful change persists over time."
- Concern : selection is not payment. Actual ACCESS model rates, covered population and renewal terms appear nowhere in the public record, and TEMPO has exactly one publicly named participant — too thin a base to conclude the route generalises to anyone else.
- Concern : the false-positive cost of a screening claim is the most under-priced item in products like this. Screening for prediabetes off continuous data funnels large numbers of asymptomatic people into the clinical system, and no positive predictive value or downstream confirmation rate appears in any of the public material — a gap worth chasing.
EFAI ERSUITE CT Appendicitis Assessment System
Emergency CT appendicitis detection · Ever Fortune.AI (Taiwan, ticker 6841)
Function and position. Deep learning reads contrast-enhanced abdominal CT and pushes a notification to clinicians when appendicitis-like features appear. It received Taiwan TFDA device approval on August 19, having already obtained US FDA 510(k) clearance in April 2026 — authorisation complete in both major markets.
- Strength : the licence count is itself the moat. The company holds 57 device authorisations worldwide — 15 US FDA, 24 Taiwan TFDA, plus Thailand, Malaysia, Vietnam and Singapore — across more than 70 customer sites. In an industry where regulatory complexity is the barrier to entry, the organisational muscle to run that process 57 times is hard to copy.
- Strength : it sits in the low-risk quadrant of the FDA's grid. Its output is a suspected-feature notification — non-directive information rather than an instruction to act — which on the activity axis of the FDA discussion paper lands at the lower-left of the grid with a correspondingly lighter burden. That is the structural advantage discriminative models still hold over generative ones.
- Concern : not a single clinical performance figure is public. Sensitivity, specificity, reduction in read time, actual effect on emergency disposition — none appears in the available sources. This is the textbook shape of the evidence deficit we keep returning to: the clearance news is complete, the outcome evidence is absent.
- Concern : 57 authorisations is a count of licences, not of revenue or installations. No pricing, paying-customer count or utilisation figure for this appendicitis product appears in the public record, and licence count must not be read as a proxy for commercialisation.
03 — Bodies & Competition
| Body / Company | Recent state & numbers | Position & moat |
|---|---|---|
| FDA CDRH / DHCoE The US device regulator |
Issued the generative AI paper on Aug 18 — 30 pages, 26 questions, comments to Oct 19; the digitally derived measures white paper on Aug 20; cumulative authorisations above 1,400 AI devices, with a record 331 in 2025. | The moat is track record and volume — no other agency on earth has reviewed this many AI devices. The weakness is that what it said this week binds no one, and the road from paper to final guidance is measured in years. |
| European Commission / AI Office Enforcer of the EU AI Act |
Reg. (EU) 2026/1744 published Jul 24, in force Jul 27: Annex III to Dec 2, 2027; embedded systems including medical devices to Aug 2, 2028; Article 50 transparency still applying from Aug 2, 2026. | The moat is market access: sell into the EU and you comply. But the deferral blunts the Brussels Effect as a template — when the strictest law in the world postpones itself by two years, other jurisdictions have less reason to copy it. |
| CMS Payment is the real gatekeeper |
The Jul 2 CY2027 OPPS proposed rule floats a SaMS category and O1 status indicator: 36 HCPCS codes, 21 into New Technology APCs, ten lab codes leaving the CLFS; it also co-runs TEMPO with the FDA. | The moat is unassailable: it sets the price. The weakness is that CMS itself calls SaMS interim valuation, and proposal-to-final can shift substantially — nobody can build long-term pricing on it yet. |
| Joint Commission + CHAI Voluntary certification and playbooks |
RUAIH launched Jun 2 across five domains, open to any organisation, and expressly not a product validation; CHAI's playbooks are the route to it. | The moat is an existing accreditation relationship hospitals already trust. The weakness is stark: it certifies an organisation's process, not a product's performance, and any buyer who reads it as a product endorsement acquires false confidence. |
| MHRA UK, the sandbox route |
AI Airlock secured £3.6m from DHSC on Apr 8 across 2026–2029, with phase three extending to LLMs, voice technologies and cancer and rare-disease diagnostics. | The moat is speed plus the NHS as a single large buyer. The weakness is scale: £3.6m spread over three years is £1.2m a year, enough to test only a handful of products — too few to generate a general standard. |
| TFDA（台灣） A follower regulator; speed is the variable |
The most recent publicly compiled list, announced January 2024, held 104 AI/ML devices — 37 domestic, 67 imported; this week it also issued the licence for Ever Fortune.AI's appendicitis CT system. | The moat is close alignment with the FDA, letting Taiwanese firms run one technical file through two markets. The weakness is that imports dominate — 67 to 37 — so domestic vendors remain a minority on home ground. |
| Dexcom Continuous glucose monitoring · US |
Named the first TEMPO selectee on Jul 22, entering the CMS ACCESS model payment pathway with its Glucose Health Program. | The moat is an installed hardware base plus long longitudinal data — something a software-only rival cannot manufacture. The weakness is fighting Abbott head-on in CGM while consumer tech companies flank into metabolic health from wearables. |
| 長佳智能 Ever Fortune.AI AI device software · Taiwan (6841) |
Took TFDA approval for the appendicitis CT system on Aug 19, bringing it to 57 authorisations (15 US FDA, 24 TFDA) across 70-plus customer sites. | The moat is industrialised regulatory throughput: maintaining a quality system across dozens of live authorisations is itself a barrier. The weakness is a broad portfolio whose individual depth is unverified in public, with no clinical performance or revenue figures available to corroborate commercialisation. |
One sentence for the section: today's competitive structure is a mismatch between binding force and clarity. The bodies with binding force — Brussels, CMS — delivered a deferral and a proposal this round. The documents with clarity — the FDA paper, CHAI's playbooks, the AI Airlock sandbox — happen to bind no one. Markets drift toward clarity, and so non-governmental bodies are quietly acquiring the substantive power to set the rules.
04 — Taiwan Angle
(1) Taiwan's dual-track clearance strategy sits exactly on the fault line this FDA paper draws. Ever Fortune.AI's appendicitis CT system took a US FDA 510(k) in April 2026 and Taiwan TFDA approval on August 19 — the island's most mature play: buy international credibility with a US clearance, then work the domestic and Southeast Asian markets. But the play rests on a 510(k) ecosystem built for discriminative models with an existing predicate. The ten benchmarking elements and competency assessment the FDA paper describes are built for generative models, and the moment a Taiwanese vendor puts an LLM into report generation, chart summarisation or intake triage, almost none of that clearance experience transfers — different tests, different evidence, different postmarket duties.
(2) The EU's Article 50 transparency duty is live now — this is a today problem, not a 2028 one. Many readers saw "high-risk deferred to 2028" and relaxed, but Article 50 was not deferred and has applied since August 2, 2026. Any Taiwanese vendor putting an interactive medical chatbot, AI-generated patient education or synthetic medical imagery in front of EU users has been under a disclosure duty for three weeks already. And the MDR/IVDR–AI Act interplay set out in MDCG 2025-6 never loosened: the real compliance weight still sits on the CE marking line.
(3) On payment, Taiwan has no SaMS of its own yet. TFDA's most recent publicly compiled list held 104 devices as of January 2024 — 37 domestic, 67 imported, so the clearance end works. The bottleneck is national health insurance. The NHIA has told the Biotechnology Strategy Advisory Committee it was evaluating AI diagnostic tools for coverage through a temporary payment mechanism, but no subsequent item list or rate schedule appears in the public record. What CMS proposed with SaMS and the O1 indicator answers exactly the same question: is an algorithmic read a separately priceable service, or an appendage to an existing procedure? If Taiwan wants its domestic AI devices to move from authorised to used at scale, it will have to answer that question itself sooner or later.
05 — Further Reading
-
FDA Releases Generative AI Medical Device Discussion Paper: What's Inside — Innolitics (2026-08)
The only piece this week that walks all ten benchmarking elements individually and flags the crucial detail that the independence requirement survives when the adjudicator is itself an LLM. Read this before deciding whether you need the 30 pages.
-
FDA Seeks Public Feedback to Inform Regulatory Approach for Generative AI-Enabled Medical Devices — FDA (2026-08-18)
The primary source, with the docket number and the deadline. If your organisation intends to comment, this is the only URL that matters before October 19.
-
EU AI Omnibus enters into force, amending the AI Act — White & Case (2026-07)
The cleanest account of what moved and what did not. Note in particular its emphasis that Article 50 was not deferred — the single most widely misread point of the week.
-
States Continue Efforts to Regulate AI in Healthcare: A Review of Legislation Passed in 2026 — Holland & Knight (2026-05-26)
Ten states sorted into four categories, with bill numbers and effective dates attached. For costing US compliance, more useful than any press release.
-
CMS floats new Medicare payment category for AI diagnostic software: 6 notes — Becker's Payer Issues (2026-07-07)
Six notes covering the SaMS mechanism and the code shifts. It explains, in about four minutes, why "FDA-authorised" and "commercially viable" are two different things in the US.
06 — References
- 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
- FDA Releases Generative AI Medical Device Discussion Paper: What's Inside. Innolitics, 2026-08. innolitics.com
- FDA Opens Comment Period on GenAI Device Regulation. MD+DI (MDDI Online), 2026-08-18. mddionline.com
- FDA issues discussion paper on regulating GenAI medical devices. AuntMinnie, 2026-08-18. auntminnie.com
- FDA digital health leader promises generative AI regulatory guidance is coming. STAT News, 2026-08-24. statnews.com
- FDA Weighs in on the Development and Use of Digitally Derived Measures for Clinical Investigations. FDA Voices, 2026-08-20. fda.gov
- TEMPO for Digital Health Devices Pilot（頁面於 2026-08-21 更新）. FDA Digital Health Center of Excellence. fda.gov
- FDA Announces First Participant Selected for TEMPO for Digital Health Devices Pilot. U.S. Food and Drug Administration, 2026-07-22. fda.gov
- Dexcom Announced as First Participant Selected for FDA's TEMPO Digital Health Devices Pilot Program. Dexcom Investor Relations, 2026-07-22. investors.dexcom.com
- EU AI Omnibus enters into force, amending the AI Act. White & Case LLP, 2026-07. whitecase.com
- EU AI Act Omnibus Agreement — Postponed High-Risk Deadlines and Other Key Changes. Gibson Dunn, 2026-05-27. gibsondunn.com
- Artificial Intelligence: Council and Parliament agree to simplify and streamline rules. Council of the European Union, 2026-05-07. consilium.europa.eu
- AI Omnibus enters into force. European Commission — Shaping Europe's digital future, 2026-07. digital-strategy.ec.europa.eu
- MDCG 2025-6: Interplay between the Medical Devices Regulation / IVDR and the AI Act. European Commission — Health, 2025. health.ec.europa.eu
- CMS floats new Medicare payment category for AI diagnostic software: 6 notes. Becker's Payer Issues, 2026-07-07. beckerspayer.com
- 7 AI health insurance state laws passed in 2026. Becker's Payer Issues, 2026-07-28. beckerspayer.com
- States Continue Efforts to Regulate AI in Healthcare: A Review of Legislation Passed in 2026. Holland & Knight, 2026-05-26. hklaw.com
- President Trump Signs Executive Order Challenging State AI Laws. Paul Hastings LLP, 2025-12-16. paulhastings.com
- Joint Commission launches voluntary AI certification program for healthcare. Fierce Healthcare, 2026-06-02. fiercehealthcare.com
- Joint Commission Releases Voluntary Responsible Use of AI in Healthcare Certification. The Joint Commission, 2026-06-01. jointcommission.org
- UK's MHRA expands AI Airlock programme with £3.6m funding boost. Open Access Government, 2026-04-08. openaccessgovernment.org
- AI in medtech is booming. Track new devices here. MedTech Dive（資料更新至 2026-03-04）. medtechdive.com
- 長佳智能闌尾炎 AI 醫療器材軟體攻台美市場，累計有 57 張海內外醫材證. 台灣光鹽生物科技學苑, 2026-08-19. biotech-edu.com
- TFDA 最新公告：核准應用 AI/ML 技術之醫療器材清單. SGS 台灣, 2024-01-22. sgs.com.tw
- 健保署揭露智慧醫療等 4 面向進展，持續推動次世代數位醫療平臺計畫. iThome, 2024-08. ithome.com.tw