1,357 to 3: the week health AI's evidence gap finally got a number
Two lines ran to their extremes this week, in opposite directions. One is deployment: Epic used its Aug 17–20 users' meeting to turn AI into an operating layer inside the EHR, claiming 85%+ of customers already use it; Stanford and peers pushed chart chatbots toward broad implementation; Rock Health counted $7.4B in H1 funding with 45% concentrated in 20 mega-deals. The other is evidence: on Aug 19 PLOS Digital Health turned the sector's validation gap into a single citable number — of 1,357 FDA-cleared AI/ML devices, just 3 (0.2%) were ever evaluated on patient outcomes such as mortality or morbidity. In the same week STAT described a shadow medical system in which 40 million Americans ask ChatGPT a health question every day — a place where even the 0.2% bar does not exist. This weekly review puts the eight things side by side.
01 — The Week in Eight
1,357 cleared, 3 validated: health AI's evidence deficit becomes a number
On Aug 19, Abulibdeh, Cajas Ordóñez, Celi (MIT) and colleagues published in PLOS Digital Health an audit of all 1,357 FDA-cleared AI/ML medical devices through Dec 5, 2025: 2.5% (34) had a registered prospective trial, 0.9% (12) posted results, 0.9% (12) reached peer review, and only 0.2% (3) evaluated patient outcomes such as mortality or morbidity. Among the 34 with trials, 73.5% enrolled fewer than 500 participants and 68% ran only in the U.S., with systematic exclusion of pregnant women, non-English speakers and other vulnerable groups.
The paper's value is not the conclusion that health AI lacks evidence — that was already consensus — but that it uses the entire clearance list as the denominator and produces a ratio you can drop straight into a policy document. Until now the evidence gap could only be argued by anecdote; from this week, any debate about reimbursement, procurement or liability for AI devices can cite 0.2%. It lands right after CMS opened an NTAP add-on payment path for AI devices and proposed a new "Software as a Medical Service" category for CY2027 — the payment machinery arrived first, the outcome evidence is still catching up.
The unit of analysis is the FDA's public AI/ML-Enabled Medical Device List as a whole; no individual vendor is named. Authors are affiliated with MIT's Laboratory for Computational Physiology and the University of Bergen, among others. Regulatory background: the FDA's official list.
Epic week: AI is promoted from feature to operating system, and the FTC walks in
Epic's Users Group Meeting ran Aug 17–20 in Verona, Wisconsin with 20,000+ attendees and shipped three things at once: Ergo, an interface layer that reassembles around the clinical context (November 2026); Agent Factory for customer-built agents (GA 2027) and Cosmos Curiosity, a predictive model trained on Cosmos data (March 2027) — with the claim that 85%+ of customers already use Epic AI and Cosmos spans 320 million people and 23 billion encounters. The same week, Reuters reported on Aug 14 that the FTC is probing Epic over non-competes and data access, with parallel reporting from STAT.
Read alongside story 1, the week's tension is complete: every number Epic put on stage is hours saved or denials avoided — not one is a clinical endpoint: 20 hours a week at ECU Health, 360+ Professional Coding Assistant implementations, 12,500 hours saved by Medical Necessity Insights. None of that is bad, but they are operating metrics, not efficacy metrics, and all are vendor- or customer-reported. And because this AI ships as a native EHR feature, it bypasses FDA device review — and therefore sits outside the denominator of that 0.2% statistic entirely.
Epic Systems (private; 43.7% acute-care EHR share); ambient voice layer powered by Microsoft Dragon; Berlin's Charité named Germany's first Epic organisation. Competitors: Oracle Health, Abridge, Commure, R1 RCM.
$7.4B across 244 deals in H1: the money did not shrink, the funnel did
Rock Health's H1 report puts U.S. digital health at $7.4B across 244 deals ($4.2B in Q1, $3.2B in Q2), $1B above H1 2025's $6.4B on 245 deals, with the median deal up to $14M — the highest since 2022. The story is concentration: 20 mega-deals across 19 companies took 45% of all capital (versus 22% in 2024), and 13 of those 19 were first-time mega-deal recipients. Examples: eMed $200M, Aidoc's $150M Series E led by Goldman Sachs (its second $150M inside a year), and $100M each for Nourish, Midi Health and Garner Health.
Forty-five percent of capital going to 20 deals means investors are paying a premium for certainty at scale, not for early innovation. That interlocks with story 1: absent clinical-outcome evidence to separate good from bad, investors fall back on observable proxies — installed base, revenue growth, logos — and those proxies structurally favour companies that are already large. Healthcare Dive reads it the same way: the room for early-stage companies is compressing while Series B and beyond stay liquid.
Aidoc (Israel; CARE clinical foundation model, deployed across 2,000+ hospitals, positioning pre-IPO); eMed, Nourish, Midi Health, Garner Health. Data: Rock Health, CB Insights.
The shadow medical system: 40 million Americans consult ChatGPT daily, under no regulator at all
Writing in STAT on Aug 19, Arya Rao and Marc Succi of Mass General argued that a full parallel system now exists outside formal care: more than 40 million Americans ask ChatGPT a health question every day, mostly outside clinic hours and with no physician involved. The cast they name: Oura selling a 50-biomarker panel through Quest Diagnostics for $99; Function Health, valued at $2.5B, offering 160 annual labs plus full-body MRI; Ro and Hims prescribing for weight loss and anxiety off asynchronous intake; and Doctronic, self-described "world's #1 AI doctor," with 24 million consultations and AI-generated prescription refills in Utah. Their own JAMA Network Open study of 21 frontier models found correct diagnosis naming above 90% with complete information, but failure to produce a comprehensive differential more than 80% of the time when given only initial-visit data.
Story 1 says AI on the clearance list lacks evidence; this one says the larger share is not on the list at all. The >90% and >80% figures are two faces of one study: models excel on cases a clinician has already curated and degrade sharply on a patient's own initial description — which is exactly the setting those 40 million people are in every day. It is the week's sharpest methodological reminder for policy: the input format of a benchmark decides whether its result generalises to real use.
OpenAI (ChatGPT), Oura × Quest Diagnostics, Function Health, Ro, Hims & Hers, Doctronic. Most operate as "health information" or telehealth and fall outside the FDA's SaMD definition.
Chart chatbots move from pilot to broad rollout — and one solved what six pathologists could not
STAT reported on Aug 20 that several large health systems are moving LLM tools for querying and summarising the EHR into broad implementation, both homegrown and vendor-built, with Stanford's ChatEHR the flagship in-house example. The case in the piece: a lymph node biopsy that six pathologists could not classify; ChatEHR surfaced evidence of sarcomatoid squamous cell carcinoma from the patient's history in a different health system and "completely explained the findings in the lymph node." The article stresses that persistent monitoring is the key to deploying these tools safely.
This is the week's one clean case in which the value of AI comes from reading what nobody has time to read, not from replacing judgement — and the use case least exposed to story 4's critique, because the input is a complete chart, precisely the regime where the JAMA Network Open study put models above 90%. Note also that it is a concrete proof of the value of interoperability: the decisive information came from another institution, so without cross-system exchange the diagnosis never appears. That thread runs straight to Taiwan's FHIR Box.
Stanford Health Care (ChatEHR, in-house); comparable commercial options include Epic's Cosmos / Agent Factory components, Abridge and Microsoft Dragon Copilot. The original is behind the STAT+ paywall; this item uses only publicly visible text.
AIDO Cell: a Nobel laureate's next move, from one protein to the whole city
On Aug 18, Palo Alto-based GenBio AI unveiled AIDO Cell, a virtual cell model. Its co-founder is University of Washington scientist David Baker, a 2024 Nobel laureate in chemistry. Where AlphaFold predicts a single protein structure, AIDO Cell aims to simulate an entire cell's molecular machinery; Baker's analogy is that AlphaFold predicts one house while a virtual cell is "predicting how that whole city works". STAT likens it to "Google Earth, but for a human cell" — zoom into proteins, RNA and DNA, then simulate a gene knockout or a drug and predict the cellular response.
The week's one technology story not directly bound by the evidence-deficit frame — a virtual cell is judged on whether wet-lab work reproduces its predictions, not on clinical endpoints. It is also the item to hold most loosely: no quantitative benchmark, validation data or funding figure was disclosed on announcement, and there is no peer-reviewed paper; STAT's write-up carries no sceptical counterpoint either. Virtual cells are one of 2026's most crowded races — until independent benchmarks land, file this under "worth tracking," not "established."
GenBio AI (Palo Alto); co-founder David Baker (UW Institute for Protein Design). Competing efforts include the Arc Institute's virtual cell programme, the Chan Zuckerberg Initiative, and the NVIDIA BioNeMo ecosystem.
Capital reshuffles at the foundation layer: Anthropic passes a $65B run rate, IPO filing reportedly near
Bloomberg and CNBC reported this week that Anthropic's Q2 revenue topped $11.5B, up more than 14-fold year over year, with a July annualised run rate above $65B, and that it could file for an IPO as soon as late August at a scale rivalling SpaceX. The same reporting has Broadcom assembling a $60–100B debt structure for AI infrastructure including Anthropic compute projects, and NVIDIA paying $6B in licensing plus a $1B investment for Poolside's technology while hiring 109 of its staff. One more item worth reading in a medical context: Google's open-weight Gemma passed 1 billion downloads with 100,000+ community variants.
Half of health AI's cost structure sits outside healthcare. As foundation-model suppliers shift from venture subsidy to debt financing and public-market pricing, two things reach hospitals: inference prices must start reflecting the cost of capital, which makes the "burn to buy share" pricing of the last two years harder to sustain; and post-IPO quarterly disclosure pushes suppliers toward high-margin, easily scaled uses — healthcare being the classic high-margin, slow-adoption case. Gemma's billion downloads point the other way: for privacy-sensitive clinical settings, on-premise open-weight models are becoming a genuine option (this week's Taiwan case is exactly that — see the Taiwan section).
Anthropic (Claude), OpenAI, Google DeepMind (Gemma), Broadcom, NVIDIA, Poolside. Note: Anthropic's financials come from Bloomberg sourcing, not company disclosure; the IPO filing timeline is reported speculation.
Europe this week: small but dense health-AI rounds, plus a fast-track that recognises other regulators
The European healthtech weekly (Aug 21) rounded up a batch of modest but AI-diagnostics-heavy deals: UK's Qureight closed a $20M Series B for AI imaging infrastructure modelling lung and cardiac disease progression; UK's TidalSense raised £16M for COPD-detection AI across Europe and the US; Netherlands' Xeltis took €20.5M, Spain's Onalabs a €9.3M Series A, Switzerland's Ahead Health €8.7M, Belgium's Azalea Vision €7.5M in EIC funding and Germany's EVERSION a €2.3M seed. On regulation: the UK MHRA published fast-track recognition pathways leaning on prior FDA, Health Canada and TGA approvals; EU notified-body reforms set maximum time caps on conformity assessment; and AI Act high-risk obligations for device-embedded AI remain deferred to 2028.
The MHRA recognition route is the week's most underrated item. It shifts regulation from duplicate review to recognising someone else's review, which for resource-constrained device-AI companies removes a large slice of the cost of going multinational — and it is precisely the model Taiwan's TFDA has been watching. Read against story 1: if every regulator moves to recognition, the 0.2% evidence deficit gets recognised along with everything else. On this pathway, regulatory efficiency and evidence quality trade off directly. Worth watching.
Qureight (Cambridge; interstitial lung disease and cardiovascular imaging AI), TidalSense (early COPD detection), Xeltis, Onalabs, Ahead Health, Azalea Vision, EVERSION; regulator MHRA. This item comes from a single weekly roundup; individual rounds are not corroborated by company press releases, so treat the figures as reported.
02 — Product Analysis
Stanford ChatEHR
EHR query & summarisation · built in-house by an academic centre · US
Function and position. Lets clinicians query and summarise a single patient's entire record in natural language, including outside records from other systems. It is positioned not as diagnostic AI but as an information-retrieval layer — turning what is already in the chart, but which nobody has time to read, into a question-answering surface (STAT, 2026-08-20).
- Strength : the input is a complete chart, which is the regime where models perform best. The JAMA Network Open work cited this week found frontier models name the diagnosis correctly over 90% of the time when information is complete — exactly ChatEHR's operating condition.
- Strength : building it in-house means the health system holds full control of prompts, model versions and audit logs, which is what makes the persistent monitoring STAT emphasises actually implementable; with a vendor product, monitoring depends on what the vendor exposes.
- Concern : the success story is an anecdote, not evidence. "Six pathologists missed it, the AI found it" is compelling, but a single case cannot answer the question that matters: how often does it miss? No sensitivity/specificity, false-positive rate or prospective trial data is public.
- Concern : the in-house route does not generalise. Stanford has an engineering team, data governance and a research ethics apparatus; most hospitals do not. If this becomes the standard, the capability gap in clinical AI widens along institutional size rather than narrowing — the same inequality that story 1 measured as trials being 68% US-only with vulnerable groups systematically excluded, seen from another angle.
GenBio AI — AIDO Cell
Virtual cell model · pre-clinical discovery · Palo Alto, US
Function and position. Simulates a whole cell's molecular machinery, zoomable to protein, RNA and DNA level, predicting cellular response to a gene knockout or a drug (STAT, 2026-08-18). It sits at the very front of discovery — not replacing wet-lab work but deciding which wet-lab work to run first.
- Strength : the evaluation loop closes. Unlike clinical AI, a virtual cell's predictions can be falsified in the wet lab within weeks. That gives it a route around the trap story 1 describes: evidence does not have to wait a decade of clinical trials.
- Strength : an unusually strong founder signal. Baker's Institute for Protein Design has a track record of moving from research to practice (RoseTTAFold among others), and the 2024 Nobel means the methodological lineage has already been validated once by the field.
- Concern : no numbers at launch. No benchmark, no validation set, no peer-reviewed paper, no disclosed raise — and STAT's piece itself carries no sceptical counterpoint. On public information alone, its distance from existing virtual-cell work cannot be assessed.
- Concern : the "Google Earth for a cell" framing hides the real question — which cell types, which species, which perturbation conditions does the training data cover? Skew in single-cell datasets across cell types and ancestries is well documented; without disclosure, virtual cells risk repeating imaging AI's familiar failure outside the training distribution.
03 — Companies & Competition
| Company | This week & numbers | Position & moat |
|---|---|---|
| Epic Systems EHR + AI platform |
UGM Aug 17–20, 20,000+ attendees; launched Ergo, Agent Factory and Cosmos Curiosity; 43.7% acute EHR share, 85%+ of customers on Epic AI, Cosmos at 320M people; under FTC antitrust probe. | Moat = installed base × Cosmos × native distribution. On the evidence spectrum: rich in operating metrics, absent in clinical endpoints — and outside the FDA's 1,357-device denominator, since none of it is a regulated device. |
| Aidoc Imaging & clinical foundation model |
$150M Series E led by Goldman Sachs, one of H1's 20 mega-deals; 2,000+ hospitals, CARE foundation model; earlier this month the first AI software to win Medicare NTAP add-on payment on a medical foundation model. | Moat = volume of clearances × a reimbursement pathway × workflow embedding. One of the few companies treating regulation-plus-payment as a moat rather than a cost — which also means that as recognition pathways spread (story 8), the moat gets shallower. |
| Stanford Health Care build-it-yourself health system |
ChatEHR moving to broad rollout, with a cross-system history solving a rare cancer classification; in parallel Cleveland Clinic published its enterprise ambient-scribe governance approach in npj Health Systems. | Moat = owning the data plus the ability to build. It is the only structural counter to Epic's bundling, but only elite academic centres can afford it. A route Taiwan's medical centres should study closely. |
| Doctronic / Function Health / Ro / Hims the shadow system |
Doctronic claims 24M consultations and AI-generated refills in Utah; Function Health valued at $2.5B with 160 annual labs; Oura sells a 50-biomarker panel via Quest for $99. | Moat = direct consumer reach plus a gap in regulatory definitions. The bottom of the evidence spectrum: neither an FDA device nor an EHR feature, so none of this week's statistics constrain it. The likeliest site of a policy event in the next two years. |
| GenBio AI virtual cell |
Unveiled AIDO Cell on Aug 18; co-founded by 2024 chemistry Nobel laureate David Baker; no benchmarks or funding figures disclosed. | Moat = talent and methodological lineage. Competitors: the Arc Institute's virtual cell effort, CZI, the NVIDIA BioNeMo ecosystem. Its advantage is a short validation loop — the wet lab can falsify a prediction in weeks, with no trial required. |
| Anthropic / OpenAI / Google foundation layer |
Anthropic at $11.5B+ in Q2 revenue and a $65B+ run rate; Broadcom raising $60–100B in debt for AI infrastructure; Google's Gemma past 1 billion downloads. | Healthcare is their customer, not their battlefield. The real transmission channels are inference pricing and model deprecation cycles. Open weights (Gemma, Qwen) are the only lever hospitals hold over their own cost and privacy. |
| Qureight / TidalSense European challengers |
Qureight's $20M Series B (lung and cardiac imaging AI) and TidalSense's £16M (COPD detection); in the same week the MHRA published recognition-based fast-track pathways built on FDA / Health Canada / TGA approvals. | Moat = disease specialisation plus access to European clinical data. Against US peers they are an order of magnitude smaller in capital, but regulatory recognition is cutting their cost of going multinational. This — not Silicon Valley — is the comparison set Taiwan's health-AI companies actually resemble. |
04 — Taiwan Angle
Taiwan's item of the week. On Aug 21, TÜV Rheinland published third-party test results for Chang-Lian Technology's patient-education assistant "Aibao": a 100% retrieval hit rate across disease, surgery, examination, medication and care-process queries; 99.9% accuracy and faithfulness in its patient-education answers; and 100% classification performance with no misclassification on out-of-scope requests, prompt-injection attempts and self-harm-related questions. The protocol references ISO/IEC TS 4213:2022 for classification-model assessment and maps to the transparency and explainability requirements of Taiwan's AI Basic Act. The system runs an on-premise self-hosted language model rather than a cloud API.
Why this small story matters against the week's international news. Story 1's core finding is "cleared but never validated." The Aibao case demonstrates another route: when a product is not a regulated device and will never reach the FDA, third-party audit can still produce a verifiable quality claim. That is exactly what is missing everywhere else — Epic's 85%, Doctronic's 24 million consultations and Function Health's $2.5B valuation have no equivalent independent audit behind them. Taiwan's population and market are too small to compete on model scale, but in the service layer of health-AI quality auditing the barrier is standard-setting capability and credibility, not GPU count. One caveat: the 99.9%/100% figures come from a single commissioned test whose question bank, sample size and difficulty distribution are not public, and they should not be equated with trial-grade evidence.
The institutional clock is running. At the Legislative Yuan's Aug 6 launch of a Smart Medical Committee, MOHW information division head Li Chien-chang said a draft of implementing rules for AI in medicine would come within three months, built on seven principles — transparency, privacy, accountability, safety, equity, sustainability and human autonomy — while the cross-system "FHIR Box" platform is slated for a year-end launch, connecting the major medical centres first and reaching roughly 80% of facilities within three to five years. Counting from Aug 6, the draft lands around November — the same month Epic ships Ergo. Three of this week's international stories map onto the three questions that draft has to answer: (1) Stanford's case in story 5 proves cross-institution records are where diagnostic value comes from, so if FHIR Box delivers retrieval without settling secondary use and AI-training consent, Taiwan gets interoperability with no usable population-evidence layer; (2) the MHRA recognition route in story 8 is a low-cost option worth TFDA's assessment, provided it is clear-eyed that recognition imports other regulators' evidence deficits along with their approvals; (3) the shadow system in story 4 exists in Taiwan too — people ask ChatGPT about their medications and lab reports — and none of the seven principles naturally covers health-information services delivered outside a licensed medical institution. Background: the MOHW's three national smart-healthcare AI centres and its 2026 launch briefing.
05 — Further Reading
-
Abulibdeh, R., et al. “1,357 AI medical devices cleared, 3 actually tested on patient outcomes” — PLOS Digital Health (2026-08-19)
The one must-read. Open access, transparent methods, and the denominator is the whole FDA list. Pay particular attention to how the methods section separates registered trials from posted results from peer-reviewed publication — that three-layer filter transfers directly to an equivalent audit of Taiwan's TFDA smart-device list.
-
Rao, A. & Succi, M. “AI's shadow medical system” — STAT First Opinion (2026-08-19)
An opinion piece by two Mass General clinician-researchers, but tightly argued and carrying its own data (the JAMA Network Open test of 21 frontier models). Read it for the policy blind spot it exposes: every existing health-AI regulatory framework assumes the AI is supplied by a medical institution.
-
“H1 2026 funding and market overview: Durable roots, shifting routes” — Rock Health (2026-08)
The cleanest market review of the week, free in full. The number to watch is not $7.4B but concentration jumping from 22% in 2024 to 45% — a shift that matters far more to an early-stage team's fundraising strategy than the headline total.
-
Trang, B. “How health systems are embracing chatbots to query and summarize patient records” — STAT (2026-08-20, STAT+)
The most complete account yet of chart chatbots moving from pilot to enterprise. Paywalled, but the visible portion already carries the key argument: the hard problem is not generation quality but monitoring after deployment. Read it with Cleveland Clinic's enterprise-deployment governance paper in npj Health Systems.
-
“This Week in European HealthTech, MedTech and Health AI” — Healthcare.digital / Nelson Advisors (2026-08-21)
A weekly European roundup with more relevance to Taiwanese readers than any US report — Europe's market scale, single-payer structures and cross-border regulatory coordination sit far closer to Taiwan's situation than Silicon Valley's does. This edition's key item is the MHRA's recognition-based fast track.
06 — References
- Abulibdeh, R., Cajas Ordóñez, S.A., Celi, L.A., Gorijavolu, R., Izath, N., Lunde, T.M. “1,357 AI medical devices cleared, 3 actually tested on patient outcomes.” PLOS Digital Health, 2026-08-19. journals.plos.org
- “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 · “Most FDA-cleared AI medical devices not tested on patient outcomes.” AuntMinnie. auntminnie.com · “Nearly All FDA-Cleared AI Medical Devices Lack Evidence of Patient Benefit.” Inside Precision Medicine. insideprecisionmedicine.com
- U.S. FDA. “Artificial Intelligence-Enabled Medical Devices” (official device list). fda.gov
- Bruce, G. “Epic's UGM 2026 preview: AI agents, Cosmos, Germany debut.” Becker's Hospital Review, 2026-08-18. beckershospitalreview.com
- “Epic unveils AI-driven Ergo Visit at 2026 UGM.” TechTarget, 2026-08. techtarget.com · “Epic expands AI ambitions with agent platform, Cosmos-powered predictions and deeper workflow automation.” Fierce Healthcare. fiercehealthcare.com
- “Epic UGM 2026 – Judy Faulkner Keynote and Cool Stuff Ahead.” Healthcare IT Today, 2026-08-19. healthcareittoday.com
- “Exclusive: FTC probes health records giant Epic Systems, sources say.” Reuters, republished by Insurance Journal, 2026-08-17 (original 2026-08-14). insurancejournal.com · Ross, C., et al. “Antitrust questions circling Epic Systems.” STAT, 2026-08-14 (STAT+). statnews.com
- Trang, B. “What Epic did — and didn't — say about AI at its annual meeting.” STAT, 2026-08-19. statnews.com
- “H1 2026 funding and market overview: Durable roots, shifting routes.” Rock Health, 2026-08. rockhealth.com · “Large funding rounds help boost digital health investment in H1.” Healthcare Dive. healthcaredive.com · “Digital health funding hits $7.4B in 2026 as AI investment reshapes the market.” Fierce Healthcare. fiercehealthcare.com
- “Aidoc Raises $150 Million Series E Led by Goldman Sachs to Scale Clinical AI for Earlier, Safer Diagnoses.” Aidoc / Goldman Sachs Asset Management. aidoc.com · am.gs.com · “Aidoc banks $150M backed by Goldman Sachs to scale clinical AI foundation model.” Fierce Healthcare. fiercehealthcare.com
- Palmer, K. “What Medicare incentives for AI-based devices mean for tech companies — and hospitals.” STAT, 2026-08-13 (STAT+). statnews.com · “STAT Health Tech: RAPID coverage for breakthrough devices, and nurses push back on AI.” STAT, 2026-08-11. statnews.com
- Rao, A., Succi, M. “AI's shadow medical system.” STAT First Opinion, 2026-08-19. statnews.com
- Trang, B. “How health systems are embracing chatbots to query and summarize patient records.” STAT, 2026-08-20 (STAT+). statnews.com
- “Accelerating ambient AI scribe enterprise-scale deployment: Cleveland Clinic's novel approach to health system-industry partnership.” npj Health Systems, 2026. doi.org/10.1038/s44401-026-00144-6
- Keshavan, M. “Prominent AI startup rolls out virtual cell model in race to speed up science.” STAT, 2026-08-18. statnews.com · genbio.ai · UW Institute for Protein Design
- “Anthropic revenue jumps to over $11.5 billion in second quarter.” CNBC, 2026-08-15. cnbc.com · “Anthropic's revenue run rate reportedly surpasses $65 billion pre-IPO.” Axios, 2026-08-17. axios.com · Fortune, 2026-08-15. fortune.com
- “Top Tech News Today, August 21, 2026: Anthropic, Apple, Broadcom, Google, Nvidia, OpenAI, Tesla & More.” Tech Startups, 2026-08-21. techstartups.com
- “This Week in European HealthTech, MedTech and Health AI: 21st August 2026.” Healthcare.digital / Nelson Advisors, 2026-08-21. healthcare.digital · qureight.com · tidalsense.com · MHRA
- 「醫療 AI 回答不能『無中生有』 德國萊因完成長聯科技『愛寶』AI 衛教測試」，新頭條 TheHubNews，2026-08-21。thehubnews.net · 同稿另見睿傳媒。right-media.news · 測試參照標準 ISO/IEC TS 4213:2022
- 「厚生會成立智慧醫療委員會 衛福部擬提AI醫療細則草案」，中央社，2026-08-06。cna.com.tw · 臺灣智慧醫療三大中心（衛福部）aicenter.mohw.gov.tw · 「衛生福利部三大AI中心啟動記者會」mohw.gov.tw
- This week's six preceding editions (for cross-reference): 8/17 臨床 · 8/18 產業 · 8/19 法規 · 8/20 技術 · 8/21 產品 · 8/22 廣義 AI