Same day, same journal, two endings: AI reading CT recovered 51 missed lesions; AI reading charts was quietly abandoned

On August 19, Rambam Health Care Campus in Israel — with collaborators at the Technion, Mayo Clinic and Beth Israel Deaconess — published a DECIDE-AI stage-1 prospective pilot in Nature Medicine. The LLM clinical decision support system SHAKED ran in one emergency-department wing while a parallel wing kept standard operations, over four weeks, with 1,138 patients analysed (584 in the intervention wing). The result splits three ways. (1) It was safe: no adverse events were detected and expert review rated 99 of 100 sampled outputs clinically appropriate. (2) The benefit did not materialise: length of stay was 4.9 hours in both wings (P = 0.99) and the 9.4-minute reduction in consultation cycle time missed significance (P = 0.077). (3) Uptake collapsed: adoption fell from 68% to 30% across the study and tracked workload closely (OR = 0.72 per shift hour, 95% CI 0.62–0.83). The one setting that bucked the trend was radiology consultation (OR = 2.98, 95% CI 1.58–5.63). The authors' conclusion is unusually blunt: "sustained clinician engagement, rather than algorithmic accuracy, may be the key barrier," and the findings "do not justify clinical deployment of AI clinical decision support at this stage."

For three years nearly every clinical-LLM study has answered one question: is the model accurate enough? SHAKED replaces the question. The model was accurate (99/100) and safe (zero adverse events), yet in a live emergency department it was used less each week — and least by the clinicians who were busiest. That inverts the commercial story about AI saving time: if AI only gets used when you are not busy, it is not addressing busyness at all. OR = 0.72 per shift hour is the coefficient of the week, because it turns adoption into something measurable and usable as an endpoint. It also annotates the PLOS Digital Health finding that only 3 of 1,357 cleared devices were tested on patient outcomes: even when you do run the trial, you may measure nothing — because the intervention was diluted by its own users mid-study. Catching exactly this is what the DECIDE-AI framework was designed for, and SHAKED is one of the few studies to follow it honestly and publish the negative result.

SHAKED is an in-house LLM decision-support system at Rambam (Haifa); authors Leibovitch, Ahituv, Gorenshtein, Aran, Sorka, Miron and Shelly are affiliated with Rambam, the Technion, Mayo Clinic and Beth Israel Deaconess. The underlying foundation model is not disclosed. The study is DECIDE-AI stage 1 (early live clinical evaluation), not an efficacy trial; the comparator is standard operations in a parallel wing of the same department.

The second Nature Medicine paper of the same day comes from Yu Shi (corresponding author) in the radiology department of Shengjing Hospital, China Medical University, with Alibaba's DAMO Academy, King's College London and EURECOM among institutions across China, France and the UK: a multicentre study plus single-arm clinical trial of the Liver DiagnOsis Network (LiON). Scale is the story: 6,443 patients for training, 22,251 for multicentre retrospective validation, 10,333 in the prospective trial. Performance: AUC 0.975 (95% CI 0.971–0.979) for malignancy in retrospective validation, 0.971 in the hepatic-steatosis subgroup and 0.924 in cirrhosis; the prospective primary endpoint was AUC 0.952 (95% CI 0.942–0.961). The weightier result is secondary: inside routine radiology workflow LiON surfaced 51 previously overlooked lesions including 15 malignancies, triggered 37 amended reports and led to 22 multidisciplinary-team escalations. It reads multiphase contrast-enhanced CT with flexible phase input (non-contrast, arterial, venous, optional delayed), outputs a patient-level malignancy score, segments eight lesion types, and folds in demographics, biomarkers and history.

LiON against SHAKED is close to a natural experiment. Both ran inside live clinical workflow, neither is a randomised efficacy trial, and only one produced identifiable patient-level consequences: 51 lesions, 37 amended reports, 22 MDT escalations. The difference is not model intelligence but the shape of the intervention. LiON exists as a second reader and safety net; its work happens outside the clinician's attention and requires no one to open anything. SHAKED asks a physician to take one more action at their busiest moment. That yields today's most usable design principle: adoption of clinical AI is inversely proportional to the active attention it demands from a clinician. The limits are equally clear: a single-arm trial has no comparator, so it cannot say whether those 51 lesions would eventually have been found without AI, nor convert them into survival; the authors themselves call for prospective comparative studies across health systems. LiON still sits outside that PLOS 0.2% — just nearer the threshold than most.

Led by the radiology department at Shengjing Hospital, China Medical University (Shenyang), with Alibaba DAMO Academy as technical partner — the same group behind the DAMO PANDA pancreatic-cancer screening model, making liver its second organ-scale target. Academic partners: King's College London and EURECOM. Competitive frame: Aidoc (abdominal CT triage), United Imaging Intelligence, Infervision, and Lunit and DeepHealth concentrated in chest and breast — multiphase liver CT is comparatively uncrowded.

Rawan Abulibdeh (University of Toronto) with co-authors including MIT's Laboratory for Computational Physiology published on August 19 in PLOS Digital Health an audit of all 1,357 FDA-authorised AI medical devices through December 5, 2025: just 34 appeared in a registered clinical trial, 12 posted results, 12 reached peer review, and only 3 were tested against patient-centred outcomes such as mortality, stroke, hospitalisation or quality of life. As News-Medical summarised on Aug 20, the research was conducted almost entirely in well-resourced health systems and systematically excluded pregnant women, adults over 75 and non-English speakers; the authors warn the framework could amplify existing disparities and make low- and middle-income countries unwitting test populations for under-studied devices. Healio (Aug 21) and AuntMinnie also covered it.

We covered this study in yesterday's weekly review; it returns today because it is the denominator for the first two stories. Place LiON and SHAKED inside the framework and something easy to miss appears: neither paper would count among the 3 — LiON is single-arm, SHAKED measured adoption and process metrics, and neither used a patient outcome as its primary endpoint. Even the two most rigorous prospective studies of the week remain some distance from proving patients live longer or better. That is not a failure of the investigators but of the evaluation infrastructure for clinical AI itself: no standard endpoints, no established convention for cross-institution comparators, no tradition of treating adoption as a formal endpoint. The structure is what to remember today, not any single number.

The unit of analysis is the whole FDA AI/ML-Enabled Medical Device List; no vendor is named. Taiwan's counterpart is the TFDA AI/ML medical device information and matchmaking platform, for which no equivalent validation audit exists.

To see why "99 of 100 appropriate" in story 1 is not itself a safety guarantee, it helps to read an earlier randomised trial. "Automation Bias in Large Language Model–Assisted Diagnostic Reasoning among Physicians Trained in AI Literacy" in NEJM AI (May 2026 issue; NCT06963957) randomised 44 physicians who had each completed 20 hours of AI-literacy training to receive either error-free or deliberately flawed LLM recommendations. Diagnostic reasoning scores fell from 84.9% to 73.3% (adjusted difference −14.0 percentage points), and top-choice diagnostic accuracy from 90.5% to 76.1% (adjusted −18.3 points). The preprint is on medRxiv, and a follow-up RCT preprint from June 2026 tests behavioural nudges as mitigation.

Nearly every risk narrative for clinical AI rests on one assumption: a physician makes the final call, so AI errors get caught. This trial shows the assumption fails even among specifically trained physicians — and the degradation (−14.0 and −18.3 points) is large enough to cancel the gains most AI tools claim. Connect it back to story 1: SHAKED's "99 of 100 appropriate" means 1 in 100 is not appropriate, and this RCT says that one will not be reliably intercepted. It is also why LiON's safety-net posture in story 2 is structurally safer: AI that says "you may have missed something" costs a second look when wrong; AI that says "here is what you should do" costs a persuaded clinician. Caveat: 44 physicians in a single country (Pakistan) in a simulated-case design — a mid-sized study whose generalisation should stay conservative.

No commercial product is involved; a general-purpose LLM supplied the recommendations, with errors deliberately injected to test independent clinical judgement. This kind of adversarial-advice stress test is not yet a pre-market requirement at any regulator — one of the gaps in the generative-AI device pathway the FDA is currently weighing.

On August 18, Healthcare IT News reported that South Korean company Ainex has deployed its real-time endoscopy AI ENAD (Endoscopy as AI-powered Device) at National University Hospital (NUH) in Singapore for upper and lower gastrointestinal endoscopy, detecting lesions in real time and automatically characterising them to support clinical decisions. Ainex says peer-reviewed studies show a 16% increase in adenoma detection rate (ADR), and that ENAD runs in more than 200 health settings in South Korea with approvals in Thailand, Malaysia and Colombia.

Endoscopy AI is the closest thing clinical AI has to a proven category: ADR is an accepted colonoscopy quality metric with a documented link to subsequent colorectal cancer incidence, and computer-aided detection (CADe) has accumulated multiple randomised trials. It is also a textbook case of story 2's design principle — the AI draws a box on the screen and no clinician has to summon it, so nothing resembling SHAKED's decay in usage occurs. Two honest annotations: (1) the 16% is a vendor-aggregated figure and the report does not link the underlying paper; (2) recent literature notes CADe raises ADR partly by removing more tiny, low-risk adenomas, leaving long-term benefit contested. The market signal is more direct: clinical-AI procurement in Korea and Singapore has entered its cross-border export phase, and Taiwanese vendors face exactly this cohort of competitors.

Ainex (South Korea); customer: National University Hospital, Singapore. Category competitors: Medtronic GI Genius, Fujifilm CAD EYE, Olympus ENDO-AID, and Korean peers. Taiwanese endoscopy-AI groups remain largely at the hospital-research-collaboration stage, with no cross-border rollout at comparable scale.

On August 21 Fierce Healthcare reported on an escalating argument: bioethicist and oncologist Ezekiel J. Emanuel, with Neal Khosla of Curai Health and Vinod Khosla of Khosla Ventures, co-authored a JAMA viewpoint arguing that after reviewing all published work on AI in medicine since January 1, 2024, AI now matches or outperforms physicians on five core cognitive tasks — gathering patient information, diagnosing, selecting tests, recommending treatment and managing chronic disease — and will be ready for autonomous deployment in some, perhaps many, workflows by around 2030. They go further: physician oversight introduces errors and degrades otherwise superior AI performance. AMA CEO John Whyte counters that medicine requires human judgement beyond task execution, likening it to pilots remaining essential in the age of autopilot, and that empathy, communication and shared decision-making cannot be delegated to an algorithm.

The debate belongs beside stories 1 and 4 because both camps can cite today's evidence, and each cites only half of it. Emanuel and Khosla can point at story 4 and say physician oversight does introduce error — the RCT does show clinicians pulled off course by bad advice. But that RCT's premise is that the AI gave the wrong answer first, which is precisely the case for not removing oversight. The AMA can point at story 1 and say clinicians do not buy it — but SHAKED's decaying uptake does not prove the AI was worthless, only that it was designed to require effort. What both sides skip is story 2's answer: arguing over whether AI should replace physicians may be the wrong level of the problem. LiON neither replaced the radiologist nor waited to be opened by one; it simply changed the probability that a lesion is seen.

Curai Health (AI-first virtual care, backed by Khosla Ventures); the American Medical Association; the University of Pennsylvania (Emanuel). Note the commercial backdrop: of the authors arguing for autonomous AI, one runs an autonomous-care company and one is its investor, so the JAMA viewpoint should be read as an interested position paper rather than a neutral literature review.

On August 22, Healthcare IT News (APAC) published an interview with Martin Seo, chief business officer at Youngand Inc., framing the structural problem in Asia-Pacific dementia care: mild cognitive impairment is the intervention window — a meta-analysis of 41 inception cohorts found roughly 39% of people with MCI progressed to dementia versus about 22% in population samples — but "detection sits in one place, delivery capacity in another, and no record comes back from either." In Singapore, roughly 2,800 private GP clinics, 27 public polyclinics and 230-plus Active Ageing Centres operate without coordination; by 2030 about one in four Singaporean citizens will be 65 or older, and dementia is projected to become the world's third leading cause of death by 2040. The company's Brain Health Playground uses motion recognition and session-level analytics to place structured cognitive training where seniors already gather.

The first six stories ask whether AI is accurate and whether anyone uses it; this one moves up a level: even a perfect screening AI changes nothing if there is no service to hand the positive case to, and no record coming back. That maps onto Taiwan closely — Taiwan is not behind on dementia screening instruments, digitised cognitive scales or imaging biomarkers, yet the path from community centres through primary care to medical centres that turns "we found it" into "they are being treated, and it is recorded" is equally missing. In story 1's vocabulary this is adoption collapse in another form: not clinicians abandoning the tool, but detected patients falling through the gaps between systems. Caveats: this is an interview with an interested vendor; Brain Health Playground has no peer-reviewed outcome evidence cited, and the 39%/22% figures come from the referenced meta-analysis, not from the company's product.

Youngand Inc. (South Korea; Brain Health Playground), operating in Singapore and across APAC. Read alongside Healthcare IT News on Aug 21 on how rural hospitals prioritise AI investment — both turn on which tier of institution the AI capability lands in.