Clinical AI posted two clean wins this week — AUC 0.913 across 146 abdominal CT findings, one extra large-vessel occlusion caught per 18 head CTs — and three other papers said our ability to tell when it is wrong has not kept pace: stigmatizing wording downgraded triage in up to 9.4% of cases, only 3 of 12 readers questioned synthetic images, and 73% of 229 digital-health trials were run by their own developers
The clinical rotation this week has an unusual shape: hard numbers landed on both the capability side and the calibration side, pointing in opposite directions. On capability, RADAR — from the First Affiliated Hospital of Zhejiang University School of Medicine and Alibaba DAMO Academy, published in Science — trained a single generalist model on 424,911 examinations and reached a mean AUC of 0.913 across 146 abdominal CT findings, 0.895 on external validation at eight centres, with the code released open-source. A Korea University model for large-vessel occlusion on non-contrast CT lifted human reader AUC from 0.718 to 0.852 — roughly one additional occlusion caught for every 18 scans read. On calibration, three npj Digital Medicine papers landed in the same week, each arguing the same thing from a different direction: we cannot tell, in the moment, when the model is wrong. Across 221,556 triage comparisons, stigmatizing wording alone produced harmful downgrades in up to 9.4% of cases. Of twelve readers, only three spontaneously questioned whether images were synthetic — and the synthetic ones were rated 0.44 higher on perceived quality than the real ones. Of 229 digital-health RCTs, 73% were run with direct developer involvement, with 23% higher odds of reporting a statistically significant result. Binding the two sides together is the FDA's September 17 final order denying a petition to exempt radiology AI from 510(k): capability can be open-sourced and spread overnight, but the premarket evidence gate is not moving.
01 — Top Stories
RADAR: one model, 146 abdominal CT findings, mean AUC 0.913 — and open-sourced
On September 17 the First Affiliated Hospital of Zhejiang University School of Medicine, Alibaba DAMO Academy and Hupan Laboratory published RADAR in Science — a vision-language model that decomposes CT volumes into individual anatomical structures and aligns them, via contrastive learning, with the corresponding text in radiology reports. It was trained on 424,911 examinations, 1.5 million image-text pairs and more than 15 million anatomy-specific pairs. Mean AUC across 146 abdominal CT findings is 0.913, against 0.776 for the best competing model; external validation across eight centres gives 0.895; and on an emergency cohort of more than 27,000 cases — a setting it was never trained for — it still reaches 0.904. Code and weights are on GitHub; the institutional release is on EurekAlert.
A decade of radiology AI commerce has been built on one algorithm doing one thing, each cleared separately. If a 146-finding generalist holds up, the unit economics of that model change: the product stops being a detector and becomes a layer of infrastructure. More consequential is the decision to open-source it — capability now spreads at a rate no single company's commercial calendar controls, while the burden of knowing whether a given version was ever validated shifts onto the clinical side. The emergency figure deserves particular attention: 0.904 on a setting the model was never trained for is exactly the kind of zero-shot generalisation claim regulators have the least machinery to assess.
Every figure here is retrospective reader-independent performance: no prospective trial, no patient-outcome endpoint, and no reader study showing what happens when a radiologist works with RADAR rather than against its benchmark. High AUC is not clinical utility. All eight external centres are in China, so generalisation across populations, scanner vendors and contrast protocols remains unproven. Open-sourcing also means no regulator stands behind whichever version ends up deployed.
Large-vessel occlusion on plain CT: reader AUC 0.718 → 0.852, one extra catch per 18 scans
Korea University College of Medicine released a Korea–US validation on September 14, published in the Journal of NeuroInterventional Surgery (DOI 10.1136/jnis-2026-025339). The model detects large-vessel occlusion (LVO) from standard non-contrast head CT alone, with no contrast agent. Across 963 patients (723 Korean, 240 US), the model scored AUC 0.963 in the Korean cohort and 0.899 in the US cohort. Eight clinical readers scored AUC 0.718 unaided and 0.852 with the model; sensitivity rose from 46.6% to 63.7% and specificity from 91.9% to 94.9%. The team frames this as one additional otherwise-missed LVO for every 18 non-contrast CTs read, and reports no significant automation bias.
This is the week's only piece of evidence that is AI-assisting-humans, in an emergency setting, with external validation across two countries — and it targets a genuine bottleneck. Non-contrast CT is the first image in most emergency stroke pathways, and unaided human sensitivity for LVO on that image sits in the forties. Lifting sensitivity into the sixties while specificity also rises makes this a rare assistance pattern that reduces misses and false alarms at once, rather than trading one for the other. Notably, the team measured automation bias directly — precisely the thing the papers below say is usually missing.
Eight readers is the study's weakest link, and the reader study is a laboratory exercise rather than the time-pressured reality of an emergency department. The "one per 18 scans" figure is derived from the sensitivity gap at this cohort's prevalence; in a setting with different LVO prevalence the number moves, so it should not be carried over to Taiwanese emergency volumes unchanged.
Dresden's on-premise clinical agent: answer consistency predicts correctness better than the model's own confidence
On September 15, TU Dresden and Dresden University Hospital published in Nature Medicine (DOI 10.1038/s41591-026-04609-x) a medical AI agent running entirely on local institutional infrastructure, tested through simulated physician-patient interactions on standardised cases. The best local model reached roughly 90% correct diagnoses on one benchmark and 84% on another; across 181 randomly selected cases, physician review agreed with the automated scoring more than 90% of the time. The central finding: whether the model gives the same answer on repeated passes predicts correctness better than its own internal probability scores — and this held under stress tests where information was deliberately removed. The study also found markedly lower accuracy on simulated cases involving older patients. Prof. Jakob Kather: "Our goal is an AI agent with selective autonomy — support clinicians in decision-making, but never take over completely."
This is the week's most immediately actionable result. The hard problem when a hospital deploys an LLM is not headline accuracy but whether to trust any particular answer — and self-reported confidence scores have a long record of poor calibration. Asking the same question three times and checking whether the answers agree is a rule any hospital can wire into a workflow the same day, with no model swap and no access to weights. Running the whole stack on-premise, with data never leaving the institution, answers the trust and the privacy procurement questions at once. The lower accuracy on older patients, meanwhile, puts a specific safety risk on the map.
Everything rests on standardised simulated cases, not real patients in a real clinic; the gap between a scripted consultation and an actual history-taking is the standing external-validity risk for this genre. The authors themselves note the cause of the age gap is unresolved — it could be data distribution, atypical presentation in older patients, or model bias, and the study cannot yet say which.
221,556 triage comparisons: adding stigmatizing wording alone pushes LLMs to downgrade patients, in up to 9.4% of cases
On September 19, the Department of Emergency Medicine and the Department of Bioinformatics at UT Southwestern Medical Center published an outcome-grounded controlled experiment in npj Digital Medicine: demographic, social and stigma-related attributes were inserted into real case presentations from two academic emergency departments, and the resulting shifts in LLM triage priority measured. The design covers 221,556 ESI-matched comparisons across eighteen neutral/stigmatizing formulation pairs, testing three open-weight models — Gemma, Qwen and DeepSeek. Framing a patient as a frequent ED user produced harmful reprioritisation in up to 9.4% of cases, running 1.4 to 7.1 points above the neutral phrasing in all six model × dataset combinations; psychiatric history produced shifts of up to 9.5%. Race, language and insurance status, by contrast, showed no consistent harmful shifts.
Most LLM bias work stops at showing that outputs differ across groups. This study anchors its endpoint in a clinical outcome — harmful downgrades after ESI matching — at two orders of magnitude greater scale, which makes the conclusion actionable in a way the genre usually is not: the risk is not abstract fairness but patient safety at the moment of triage. The counterintuitive part matters most to hospitals. Race and insurance showed no consistent effect; what actually harmed patients was wording clinicians themselves put in the chart ("frequent flyer", "history of psychiatric illness"). De-identification alone will not protect against this; the narrative itself has to be governed.
Only three open-weight models were tested, so the findings do not transfer directly to the closed commercial models most hospitals actually run; the 9.5% psychiatric-history shift reached statistical significance in only one model. And this is simulated triage ordering, not outcome tracking after real deployment — it establishes that the risk exists, not that harm has occurred.
73% of 229 digital-health RCTs were run by their own developers, with 23% higher odds of a significant result
On September 17, npj Digital Medicine published "Clinical trials for digital health interventions: a rapid review of study independence and the developer effect", extracting 229 RCTs from 29 systematic reviews spanning 2013–2025 across nutrition, maternal health, mental health and sleep. 73% of trials involved the developer directly; only 27% were independent. Developer-involved trials were markedly more likely to be preregistered (OR 2.47, p=0.004) — procedurally, they were the tidier ones. But weighted by sample size, they also had higher odds of reporting a statistically significant result: OR 1.23 (95% CI 1.16–1.31, p<0.001).
This gives a discount factor that can be applied directly to any clinical AI result. The theme this site has worked through for two weeks — vendor-stated benefit running three to six times higher than peer-reviewed benefit — gets a mechanism here: the problem is not fabrication but that trials designed, run and analysed by the developer are systematically more likely to produce significant findings. That developer-involved trials were the more likely to be preregistered is the sharpest detail: procedural compliance is not a proxy for independence. The first question to ask of any clinical AI paper is therefore whether anyone on the author list sells the thing.
The sample is digital-health interventions, behavioural apps included, not purely clinical AI device trials, so extrapolating to imaging AI or clinical decision support needs care. An OR of 1.23 is a statistically robust but modest shift — enough to adjust a prior when reading the literature, not enough to overturn any single study.
Only 3 of 12 readers questioned the images — and the synthetic breast ultrasounds looked better than the real ones
On September 19, npj Digital Medicine published a three-phase staged evaluation. In a purpose-blinded first phase, twelve readers each assessed 60 breast ultrasound images per source without being told any were synthetic. Reader-perceived quality came out higher for the synthetic images, by 0.44 (95% CI 0.33–0.55); model-estimated sufficiency for basic assessment was 95.9% for synthetic versus 86.7% for authentic. Only three of the twelve readers spontaneously raised a question about provenance. Told explicitly that synthetic images were present, readers detected them with 70.8% accuracy; given identification cues, 79.5%. Ultimately, in 94.9% of synthetic-image evaluations readers chose to review, verify or not use the image.
Synthetic medical imagery has been pitched as a data-augmentation fix — not enough rare lesions, so generate some. This study reframes it as a risk. When generated images are perceived as better quality than real ones, the provenance chain running through training sets, teaching files, teleradiology pipelines and research databases becomes a clinical safety concern, and no hospital process currently checks it. The number that stings is 3 of 12: readers are not incapable of spotting synthetic images — told to look, they hit 70%. They simply never think to ask. That pairs directly with the Dresden result: the calibration failure is not in the model, it is that nobody gave the human a moment to ask the question.
Twelve readers, a single modality: a very small sample. The authors state explicitly that the work does not establish clinical fidelity, diagnostic equivalence or causal intervention efficacy for synthetic images. What it demonstrates is that humans do not spontaneously question provenance — not that synthetic images cause misdiagnosis.
AI-assisted colposcopy, externally validated: +6.5 points of sensitivity for junior clinicians, biopsies down from 2.48 to 2.02 per case
On September 19, npj Digital Medicine published an independent external validation on the WHO open-source dataset: 187 patients, 855 colposcopic images, read by 45 colposcopists of varying experience across 12 regions of China. AI alone reached 84.2% sensitivity for CIN2+ (95% CI 74.4–90.7); clinicians unaided scored 84.8% (95% CI 82.7–86.9) and 90.6% with AI (95% CI 88.7–92.1). For low-experience readers, AUC rose from 0.72 to 0.76 (p=0.043), a 6.5-point sensitivity gain. Mean biopsies per case fell from 2.48 to 2.02.
Two things. First, this is external validation on somebody else's public dataset rather than a vendor's own held-out split — which, in the light of the developer-effect paper above, carries a different evidential weight entirely. Second, it improved two metrics that usually trade against each other: sensitivity up, biopsies down. The standard cost of assistive diagnostic AI is overtesting; here invasive procedures fell instead, which matters most in settings where cervical screening capacity is scarce. And the gain concentrates in junior readers — exactly the workforce profile of those settings.
187 patients is a small validation set and the CIN2+ confidence interval is wide (74.4–90.7 for AI alone); an AUC move from 0.72 to 0.76 is clinically marginal and p=0.043 sits close to the threshold. All readers were in China, where colposcopy training differs from Taiwan's. Fewer biopsies is welcome, but the study did not track whether any lesions were missed as a result.
FDA denies the radiology-AI 510(k) exemption petition: every clinically meaningful change still needs a new submission
On September 17 the FDA published a final order in the Federal Register (Docket FDA-2025-P-5560) denying a petition filed on October 22, 2025 by Nancy Stade of Rubrum Advising on behalf of Harrison.ai. The petition sought partial exemption from 510(k) for radiology computer-aided detection/diagnosis (CADe/CADx) and computer-aided triage and notification (CADt) software, conditional on the manufacturer holding prior clearances and operating robust post-market surveillance. Applying its four exemption factors — device history, performance characteristics, detectability of changes and classification stability — the agency found none satisfied, stating that "the information presented in the petition does not demonstrate that premarket notification is not necessary," while reaffirming a commitment to "innovative and least burdensome" digital health pathways. Full order in the Federal Register; industry read-through at PYMNTS.
Set beside the week's other seven items, this one gains weight. In the same week that RADAR open-sourced a 146-finding generalist, the FDA told Harrison.ai — a company with multiple clearances deployed across more than 1,000 facilities — that its next version still has to queue. The agency is answering a narrow, concrete question: can post-market surveillance detect the performance change introduced by a model update? Its answer is no. That is precisely what the two September 19 papers support empirically — humans do not spontaneously question, confidence scores are not trustworthy — so the gate has to stay upstream. The commercial consequence is that revenue is pinned to versions: which build carries clearance has to be traceable.
This denies one petition; it is not a new policy statement, and it does not disturb the existing predetermined change control plan (PCCP) pathway, through which manufacturers can still obtain advance authorisation for a defined envelope of updates. Reading it as a general FDA tightening overstates it. The Federal Register body text could not be retrieved directly in this environment; document identifiers and the quoted language were cross-checked between the register's own page and the PYMNTS report.
CHOP builds its own cardiac modelling on open-source MONAI: four hours down to seconds, about 200 cases this year
On September 15, NVIDIA's blog described the Children's Hospital of Philadelphia's cardiac modelling service: using MONAI, MONAI Label, NVIDIA Auto3DSeg, 3D Slicer and the SlicerHeart extension, it generates anatomically precise heart models from existing CT, MRI and 3D ultrasound for congenital heart surgery planning. Per-case modelling time falls from four hours to seconds. CHOP expects to model roughly 200 cases this year; more than 20 US children's hospitals run modelling programmes, and Boston Children's supports about 500 surgeries a year this way. Congenital heart defects affect roughly 1% of live births. CHOP cardiologist Matthew Jolley: "You've got a one-of-a-kind kid and an off-the-shelf device. Our job is to find what fits."
This is the week's only case of a hospital building its own tool, running it in its own workflow, with a quantified improvement — and the architecture matches RADAR and Dresden: open-source components, deployed on-premise. Four hours to seconds is not merely time saved; it changes whether modelling can enter routine pre-operative practice at all. At four hours it is worth doing only for the most complex cases; at seconds it can be done for every operation. Note too that this route sidesteps the commercial device clearance cycle entirely, because it is surgical planning support rather than a diagnostic claim.
This is NVIDIA's own blog, not a peer-reviewed study, and nothing here is independently audited. "Four hours to seconds" is process time, not a clinical outcome — there are no surgical success, complication or mortality figures in the piece. The ~200 cases and ~500 surgeries are institution-reported.
02 — Product Analysis
RADAR（damo-radar）
Generalist abdominal CT vision-language model · Zhejiang University First Affiliated Hospital + Alibaba DAMO Academy (China)
Function and position. It decomposes contrast-enhanced abdominal CT volumes into individual anatomical structures, then aligns them with radiology report text through contrastive learning, yielding a single model that reads out 146 findings. The positioning is not a detector for one specialty but an interpretation layer any downstream application can call, with the paper and code released together.
- Strength : breadth and generalisation hold at once — mean AUC 0.913 across 146 findings against 0.776 for the best competing model, 0.895 on eight-centre external validation, and 0.904 on an untrained-for emergency cohort of 27,000-plus cases. Zero-shot transfer to an unseen setting is the hardest capability claim for a generalist model to fake.
- Strength : open weights change the adoption curve directly. Any hospital or startup can rerun the evaluation on its own data and deploy locally, without waiting for a vendor licence and without sending images off-premise — the same architecture Dresden and CHOP independently chose this week.
- Concern : the one number that matters most is absent — there is no reader study. Every figure measures the model working alone, while the deployed form is always radiologist-plus-model. This is exactly why the Korea University result is closer to usable: it measured the human-plus-AI delta, 0.718 → 0.852.
- Concern : no regulatory path and no version governance. The FDA restated in the same week that every clinically meaningful change requires resubmission. Once an open model is forked, fine-tuned and updated inside a hospital, nothing tracks which build is live or whether that build was ever validated. All eight external centres are in China, so cross-population and cross-vendor generalisation is unproven.
TU Dresden on-premise clinical AI agent
On-premise diagnostic conversational agent · TU Dresden + Dresden University Hospital (Germany)
Function and position. Two agents that work standardised cases through simulated physician-patient dialogue, running entirely on the hospital's own infrastructure with data never leaving the institution. The best local model reached roughly 90% correct diagnoses on one benchmark and 84% on another — but the team's claim is not about the score. It is about the criterion for when the thing can be trusted. Prof. Jakob Kather calls the target "selective autonomy".
- Strength : it delivers a trust rule that can be implemented the same day — repeated-answer consistency predicts correctness better than the model's self-reported confidence, and holds under stress tests with information deliberately removed. It needs no weights and no retraining, so a hospital on a closed API can apply it too: ask three times, escalate to a human when the answers disagree.
- Strength : it publishes its own failure mode. The study reports markedly lower accuracy on older patients — the kind of disclosure that almost never appears in vendor-led work. Against this week's finding that 73% of trials are developer-run with 23% higher odds of a significant result, that difference in evidential weight is the whole point.
- Concern : everything rests on standardised simulated cases — no real patients, none of the time pressure or incomplete information of an actual clinic. The gap between scripted consultation and real history-taking is the standing external-validity risk of this genre, and the study offers no prospective deployment data at all.
- Concern : consistency as a trust signal has a structural weakness — a model can be reliably, consistently wrong. The study shows consistency predicts correctness better than confidence, but does not report the residual rate of consistent-and-wrong answers, and that figure is what determines whether it can serve as a safety threshold. The authors also state plainly that the cause of the age gap is unresolved.
03 — Companies & Competition
| Company | Recent state & numbers | Position & moat |
|---|---|---|
| Alibaba DAMO Academy Alibaba's research arm, with Zhejiang University's First Affiliated Hospital |
Published RADAR in Science on 9/17: 424,911 training examinations, mean AUC 0.913 across 146 abdominal CT findings, 0.895 on eight-centre external validation, 0.904 on a 27,000-plus emergency cohort — and open-sourced. | The moat is data scale and a top-journal imprimatur, not a product. Open-sourcing forfeits licensing revenue in exchange for ecosystem position — letting other people's products grow on its model. The weakness is no regulatory path, no reader study and no commercial channel: a hospital adopting it owns the validation and version governance itself. |
| Harrison.ai / Annalise.ai Australian radiology AI · flagship Annalise Enterprise CXR |
Closed a US$112M Series C led by Aware Super in February 2025, US$240M raised to date; 12 FDA clearances, deployed at over 1,000 facilities, used by half of Australia's radiologists, touching 6 million patients a year. On 9/17 its 510(k) exemption petition was denied by the FDA. | Its moat is exactly the layer RADAR lacks: clearances, reimbursement (one brain CT algorithm carries an NTAP) and an installed base. But Annalise CXR's claim of up to 124 chest X-ray findings, set against a free 146-finding abdominal CT model, shows breadth of coverage shifting from a differentiator to table stakes. Losing the petition is an admission that its cost structure stays pinned to per-version review. |
| NVIDIA（MONAI 生態） Supplier of the substrate layer for medical imaging AI |
On 9/15 it showcased CHOP building its own cardiac modelling on MONAI, Auto3DSeg and 3D Slicer: four hours per case down to seconds, ~200 CHOP cases this year, 20-plus US children's hospitals running modelling programmes, ~500 surgeries a year supported at Boston Children's. | It sells no device and files nothing with the FDA; it sells the layer everyone building their own thing has to use. The more hospitals self-build, the safer its position — and RADAR, Dresden and CHOP all independently chose open-source self-build this week, all landing inside its moat. The weakness is that this layer's value cannot be demonstrated in clinical outcomes; the numbers are institution-reported process times. |
| TU Dresden / Dresden University Hospital An academic institution defining a product category |
Published its on-premise diagnostic agent in Nature Medicine on 9/15 (DOI 10.1038/s41591-026-04609-x): ~90% and 84% accuracy on two benchmarks, over 90% physician-review agreement across 181 cases, consistency beating confidence scores, and lower accuracy on older patients. | No business model, yet arguably the week's biggest threat to vendors: if on-premise self-build with open weights and a consistency threshold works, one of the necessary justifications for clinical decision support licence fees disappears. Its moat is methodology rather than code — the "when to trust it" criterion can be copied by anyone, vendors included. |
| Google Health One of the main forces behind conversational medical AI |
Published a comment in Nature Medicine on 9/14, co-authored with Beth Israel and others, arguing that trust in conversational medical AI cannot be built on benchmarks and requires prospective real-world studies. It is a comment piece with no quantitative results. | The party that has led on benchmarks longest is arguing benchmarks are insufficient — which moves the bar toward what it can afford, since only large players can fund prospective trials. That squeezes smaller vendors and helps hospitals: it points to the same conclusion as the Dresden paper and the developer-effect review this week, from a different motive. |
Today's competitive structure has the shape of free capability and commodified trust. RADAR turns coverage breadth into a public good, while the denial of Harrison.ai's petition confirms that clearance and reimbursement remain the scarce resources — so the moat is migrating from how accurate the model is to who can prove it is accurate on your patients. Dresden's consistency criterion, Korea University's reader-study delta and the colposcopy team's external validation on a WHO public dataset are all the same asset: a validation method a third party can reproduce. The fact that 73% of trials are still run by the developers themselves shows how scarce that asset currently is.
04 — Taiwan Angle
(1) The health ministry is betting on data standards, and the timing is right. According to a September 15 report in United Daily News, Minister of Health and Welfare Shih Chung-liang argued that the first task in making medical AI land is "establishing data standards, a governance regime and a trustworthy usage environment", with FHIR Box deployments at medical centres handling standardised conversion, key clinical data connected across all national medical centres by year end and all hospitals by the end of next year. The attached resources include NT$48.9bn over five years for the Healthy Taiwan deepening programme, NT$24bn over four years for national pharmaceutical resilience, and NT$10bn in strategic investment. Read against this week's RADAR release, the sequencing makes sense: once world-class models are free, Taiwan's bottleneck is not whether it can buy a model but whether it has usable, linkable, consistently labelled local data to validate one.
(2) But what Taiwan should copy this week is the validation method, not the model. All eight of RADAR's external validation centres are in China and there is no reader study; adopting it directly means accepting a performance claim never tested on a local population. The three genuinely reproducible methods this week are Korea University's reader-AUC delta with and without AI (0.718 → 0.852), Dresden's use of repeated-answer consistency as a trust threshold, and the colposcopy team's external validation on somebody else's public dataset. None requires building a model or buying compute at scale — only a group of clinicians willing to run a reader study, and clean local data. The latter is precisely what FHIR Box is meant to deliver.
(3) Taiwanese emergency departments should read the stigma paper now. UT Southwestern's 221,556 comparisons show that harmful downgrades came not from race or insurance status but from wording clinicians put in the chart themselves — "frequent flyer", "history of psychiatric illness". Taiwan has the same populations of high-frequency attenders and patients with psychiatric comorbidity, and the shorthand and idiom of Chinese-language charting — 老病人 ("old regular"), 情緒個案 ("emotional case") — has never been tested by any bias study. Before an LLM is wired into triage, this is the local pre-check to run, and it is cheap: it needs only matched cases and a consistent severity baseline.
05 — Further Reading
-
RADAR: an expert-level generalist AI for abdominal CT diagnosis — Science (2026-09-17)
Read the methods section alone: how it decomposes CT volumes into anatomical structures and aligns them with report text is the one architectural idea this week that repays close reading. The accompanying GitHub repo lets you rerun the evaluation on your own data.
-
Clinical trials for digital health interventions: study independence and the developer effect — npj Digital Medicine (2026-09-17)
The most practical correction table published this year. Afterwards, the first thing you do with any clinical AI paper is turn to the conflict-of-interest statement rather than the primary endpoint.
-
Outcome-grounded effect of clinically stigmatizing information on LLM emergency triage prioritization — npj Digital Medicine (2026-09-19)
Read it for the experimental design rather than the finding: anchoring the bias endpoint in harmful downgrades after ESI matching is the cleanest reproducible protocol available, and any hospital can run it against its own charting corpus.
-
Provenance risk of synthetic breast ultrasound images: a three-phase staged evaluation — npj Digital Medicine (2026-09-19)
A small sample, but it asks a question nobody else is asking yet: when synthetic images are perceived as better than real ones, who is checking provenance in your teaching files, research databases and teleradiology pipelines?
-
Prospective evidence for conversational medical AI is non-negotiable — Nature Medicine Comment (2026-09-14)
No data, but worth reading for who is saying it: a Google Health and Beth Israel team arguing that benchmarks cannot build trust. The identity of the author list is as informative as the argument.
06 — References
- RADAR: an expert-level generalist AI for abdominal CT diagnosis. Science, 2026-09-17. science.org · PubMed 42752131
- Introducing RADAR (press release). EurekAlert / AAAS, 2026-09-17. eurekalert.org
- damo-radar (source code and weights). GitHub — Alibaba DAMO Academy, 2026-09. github.com
- AI detects large vessel occlusion on non-contrast CT: Korea–US validation. Korea University College of Medicine / Journal of NeuroInterventional Surgery, 2026-09-14. DOI 10.1136/jnis-2026-025339. eurekalert.org
- On-premise medical AI agents with selective autonomy. TU Dresden / Nature Medicine, 2026-09-15. DOI 10.1038/s41591-026-04609-x. eurekalert.org
- Outcome-grounded effect of clinically stigmatizing information on large language model emergency triage prioritization. npj Digital Medicine, 2026-09-19. nature.com
- Clinical trials for digital health interventions: a rapid review of study independence and the developer effect. npj Digital Medicine, 2026-09-17. nature.com
- Provenance risk of synthetic breast ultrasound images: a three-phase staged evaluation. npj Digital Medicine, 2026-09-19. nature.com
- External validation of AI assisted colposcopy using WHO dataset for cervical precancer and cancer detection. npj Digital Medicine, 2026-09-19. nature.com
- Medical Devices; Exemption From Premarket Notification: Radiology Computer-Aided Detection and/or Diagnosis Devices (final order, Docket FDA-2025-P-5560). Federal Register 91 FR 58817, 2026-09-17. federalregister.gov
- FDA keeps radiology AI revenue tied to premarket clearance. PYMNTS, 2026-09-17. pymnts.com
- Children's hospital turns to open-source AI for cardiac care. NVIDIA Blog, 2026-09-15. blogs.nvidia.com
- Prospective evidence for conversational medical AI is non-negotiable (Comment). Nature Medicine, 2026-09-14. DOI 10.1038/s41591-026-04639-5. nature.com
- 石崇良：醫療 AI 落地首重資料標準與治理制度. 聯合新聞網（經光鹽生物科技學苑轉載）, 2026-09-15. biotech-edu.com
- Radiology AI firm Harrison.ai raises $112M, opens US office. Radiology Business, 2025-02-11. radiologybusiness.com