Longevity InTime: Autonomous AI Institute. Anti-Aging Digital Health Immortality Transhumanist AI Channel

C5R’s AI‑run lab shows top model solves less than half of real science tasks C5R built Facility‑0, an AI‑managed laboratory, in 12 weeks. The system selects the next experiment, operates some instruments directly, and gives step‑by‑step instructions to human technicians for the remaining steps. A model receives a scientific goal — such as synthesizing a compound or reading a spectrum — then studies inventory, designs an experiment in code, translates that code into equipment commands and human‑readable directions, and after each run evaluates the results to decide what to try next. On 24 September the company released SciUniverse Level 1, a benchmark of 92 tasks spanning chemistry, biology and materials science — from organic synthesis and PCR DNA amplification to pressing ceramic tablets. The leading model, Claude Fable 5.1, succeeded on 45.3% of attempts on the first try, while the other five tested models scored lower. During testing the model controls instruments, lab staff and the reagent warehouse as a single agent, a fact confirmed by an independent audit of the system. After each experiment it reads measurements, plans the next move while respecting limited reagent supplies and prior outcomes. C5R says earlier AI‑in‑science evaluations only examined data after experiments were done, missing the full workflow of material selection, instrument operation and troubleshooting. SciUniverse closes that gap by measuring the entire chain from goal to outcome, including both physical runs in Facility‑0 and simulations on a digital twin of the lab using real experimental data. Models generally handle experimental strategy and calculations but often fail on manual details: pipetting frozen samples, vortexing open well plates, reusing a single pipette tip for multiple DNA wells (cross‑contaminating samples), and ignoring solvent evaporation during long reactions. Founder Michael Akilian previously worked on wearable electronics at Apple and Misfit Wearables, then founded and sold the AI‑scheduling startup Clara Labs. About eighteen months ago he shifted to experimental biology, building a worm lab in his apartment before joining a UC San Francisco laboratory, saying “Models succeeded in the virtual world but fell 🔗 Read original →

Second Lawsuit Challenges FDA Over Aging-as-Disease Petition This is the second lawsuit filed against the FDA regarding its demand to recognize aging as a disease. The first suit, filed about a year ago, was dismissed on a technicality—the missing signature—without addressing the substantive issue. Now a scheduling order has been issued. The core complaint is simple: the agency is ignoring a civil petition to classify aging as a disease that was submitted on 22 May 2024, demonstrating silence for more than two years. Let’s be clear: the odds that the FDA will tomorrow, under judicial pressure, concede and label aging a pathology are essentially zero. A Federal court lacks the authority to make medical decisions for a specialized agency. Under the Administrative Procedure Act, the most Afshar can achieve is to compel the bureaucratic machine to issue some official response—whether a denial, an agreement, or a defined timeline. Most likely, the FDA will simply reply with a boilerplate statement about “multifactorial natural processes,” thereby closing the procedural matter. Nonetheless, such lawsuits have concrete value. Each procedural clash, even if lost, generates informational noise, forces the system to react, and gradually shifts the Overton window. By repeatedly creating precedents and documenting regulatory obstruction in the public sphere, transhumanists turn radical life extension from a sci‑fi fringe topic into an inevitable public discussion. 🔗 Source: @solid_state_humanity

How are you my friends? We are organizing the International Longevity Summit in Madrid, and please join us virtually, if you can't come in person! It will be immortal, very literally, so share with your friends too, please 🖖 https://longevitysummit.madrid/

Mitochondrial transplant reverses heart aging in mice via HIF‑3α/BNIP3 switch Researchers at Shanxi Medical University in China transplanted healthy mitochondria into the hearts of aged mice and found that the procedure restored cellular energy and pump function. They identified a molecular switch that drives the breakdown of mitophagy in aging hearts. The findings were published Aging Cell, 23 September 2026. In older hearts, the protein HIF‑3α shifts to a full‑length form that activates BNIP3, causing an overload of damaged mitochondria that cannot be cleared 🔗 Read original →

This is the second lawsuit filed against the FDA regarding the requirement to recognize aging as a disease. Officials dismissed the first lawsuit a year ago due to a simple lack of signature, without ever getting to the point, but now a subpoena has been issued. The gist of the complaint is simple: the agency is ignoring a citizen's petition to recognize aging as a disease, dated May 22, 2024, and has demonstratively remained silent for over two years. Let's get this out of the way: the chances that the FDA will cave under the judge's hammer tomorrow and declare aging a pathology are slim to none. A federal court has no authority to make medical decisions for the agency in question. The only thing Afshar can achieve under the Administrative Procedure Act is to force the bureaucratic machine to issue some kind of official response: a refusal, an agreement, or a specific deadline. Most likely, the FDA will simply issue a boilerplate statement about "multifactorial natural processes," closing the procedural matter. But such lawsuits have a concrete benefit. Every such legal battle, even one lost procedurally, creates information noise, forces the system to react, and gradually shifts the Overton window. We can't breach this wall in a single blow, but by constantly setting precedents and documenting regulatory sabotage in the public arena, transhumanists are transforming the topic of radical life extension from a fantasy fantasy into an inevitable public debate.

Frozen trial blood hints that drugs may already be slowing aging On September 24, biological‑age researcher Christine Gloriozo published an analysis showing that most drug trials collect and store participants’ blood but rarely test whether the treatment slowed aging. She examined two cases where such a check was done retrospectively — semaglutide and the experimental antifibrotic drug rentoceritib (a TNIK inhibitor from Insilico Medicine) — and found that aging clocks indicated a slowdown in treated patients. In the semaglutide trial of 108 participants with HIV, 84 retained samples were available for aging‑clock analysis. Using 17 clocks, the PhenoAge score showed the semaglutide group was on average 4.9 years younger than placebo, while DunedinPACE indicated a 9% slowing of aging pace. The effect was not uniform across all clocks, but seven of eleven organ‑system clocks showed significant slowing, including brain, heart, kidney and liver — a pattern unlikely to be random noise. Among the seven paper authors, two are employees and one is a consultant of TruDiagnostic, the developer of the epigenetic clocks used. The second example Gloriozo cites is rentoceritib, where six proteomic clocks from archived Phase IIa serum consistently showed reduced biological age in treated patients. She notes that publishing null results is also important: in the RAPA‑EX‑01 trial (rapamycin plus exercise) none of four clocks showed significant slowing. Separately, another study applied protein‑based aging clocks to five semaglutide trials involving 10,052 participants and observed the same slowing effect. Re‑analyzing stored blood from completed trials is a low‑cost way to build evidence that regulators might eventually accept aging‑clock slowing as a basis for drug approval. Each uns analyzed trial in a biobank represents a paid‑for but unasked question about aging. 🔗 Read original →

Gene therapy delivering Hsp70 extends lifespan in ALS‑model mice Russian researchers administered a single injection of an AAV vector carrying the human Hsp70 gene to mice with an ALS‑like condition caused by a FUS mutation. At the high dose, median survival increased from 128 to 170.5 days and the symptomatic stage was lengthened by 30% (18.7 days vs 14.4 days). Treatment began at day 58–60, when aggregated FUS had already accumulated in motor neurons, so the onset age of symptoms did not shift. Average lifespan rose by 23% at high dose, and the AAV vector remained active in brain and spinal cord for at least 150 days. Hsp70 works as a chaperone that refolds misfolded proteins and dissolves toxic aggregates; mutant FUS fails to fold correctly and forms clumps that overwhelm the cellular quality‑control system. Hsp70 levels fall with age and neurodegeneration. The preprint, 21 September showed that the 🔗 Read original →

Body‑Conductive Network Enables Millimeter‑Scale Implants for Health Monitoring Traditional radiofrequency standards such as Bluetooth and NFC penetrate biological fluids poorly and require bulky antennas and large batteries. The new approach uses the body itself as a conductor, sending ultra‑short electrical pulses that do not damage surrounding tissue, allowing distributed sensors and actuators to operate as a single autonomous network. By eliminating radiofrequency modules, implant dimensions shrink to less than 3 millimeters, enabling delivery deep into tissues or the stomach with a standard syringe and no major surgery. Each device consists of passive electronic components that consume virtually no power in standby and activate instantly upon receiving an impulse of specific voltage and duration. In experiments, a microactuator triggered once per day operated for about one year on an autonomous battery. The system, dubbed SWANS, was demonstrated on a rat locomotion‑synchronization model where a sensor detecting fore‑limb movement sent a signal through bodily tissues to an implant in the hind limb, eliciting precise muscle contraction and mimicking natural gait. The SWANS network is designed to transmit only minimal data—key triggers and activation commands—while complex computations are offloaded to an external wearable hub. 🔗 Read original →

Pancreatic alpha cells switch to insulin production when m6A lock is removed On 18 September 2026, researchers at the Joslin Diabetes Center of Harvard Medical School reported that the enzyme METTL14 maintains the specialized state of pancreatic alpha cells. When the METTL14 gene was deleted specifically in alpha cells of mice, some of those cells began to produce insulin instead of glucagon. The team traced the effect to a single nucleotide in the RNA of the YY1 gene, showing that the same regulatory node operates in human pancreatic cells. In the islet, beta cells secrete insulin to lower blood glucose, while alpha cells secrete glucagon to raise it. Earlier work in 2010 by Pedro Herrera’s laboratory in Geneva had demonstrated that after near‑total loss of beta cells, mouse alpha cells can start making insulin, but the factor that prevents this conversion in a healthy organ was unknown. The new study identified that brake: the METTL3/METTL14 complex deposits the m6A chemical mark on alpha‑cell RNA, and the level of this mark fluctuates with nutritional cues. Amino acids such as L‑arginine and low glucose stimulate glucagon release and, in a dose‑dependent manner, increase METTL14 and m6A; insulin and palmitate reduce the mark, whereas METTL3 showed little response in these experiments, focusing further work on METTL14. Deleting Mettl14 only in mouse alpha cells abolished glucagon secretion in response to amino acids, reduced alpha‑cell mass, and increased beta‑cell mass. The newly insulin‑producing cells retained an alpha‑cell lineage label, indicating they arose by direct conversion in situ, without transplantation or organ injury. A parallel experiment targeting the ALDH3B2 gene in pancreatic duct cells activated the insulin gene only after those cells were cultured outside the body and transplanted under the kidney capsule of diabetic mice. Single‑nucleus sequencing revealed that the conversion is incomplete: the cells lose alpha‑cell genes, gain insulin‑related genes, but retain immaturity markers, which the authors describe as progression toward beta identity rather than a finished state. By mapping m6A‑marked RNA regions against factors altered upon METTL14 loss, the authors pinpointed YY1 as the sole m6A target. Without the m6A mark, the YY1 transcript is degraded more slowly, leading to protein accumulation; a point mutation at the modified nucleotide confirmed this mechanism directly. Elevated YY1 suppresses alpha‑cell identity genes PAX6 and MAFB and activates growth pathways similar to those seen with METTL14 loss; both excess and deficiency of YY1 disrupt identity, showing it acts as a dosage‑sensitive regulator, not a simple switch. Analysis of 🔗 Read original →

CerS6 enzyme drives kidney podocyte aging via C16 ceramide–p53 pathway Scientists identified the enzyme CerS6 as a trigger of aging in kidney podocytes, the cells that form the blood‑filtering barrier, and showed that disabling it in mice slowed age‑related kidney damage. The findings were posted as a preprint on Research Square, 21 September 2026 by a team from the University of Texas Health Science Center at San Antonio, which demonstrated that CerS6 in podocytes promotes aging through the accumulation of C16 ceramide and activation of the p53 protein. In aged marmosets (~16 years) the CerS6–C16 ceramide–p53 pathway was more active than in young animals (~3 years), and mice with podocyte‑specific CerS6 knockout exhibited markedly weaker signs of kidney aging. Podocytes wrap the glomerular capillaries and do not divide, so loss of these cells irreversibly weakens the filtration barrier. Chronic kidney disease affects 38% of people over 65, and podocyte aging is one of its mechanisms. The researchers studied marmosets—New World primates whose genome was sequenced in 2014—as an aging model. Aged animals showed more senescent cells in glomeruli, higher p53/p21 levels, increased CerS6 and C16 ceramide in podocytes, and a correlation between higher CerS6, greater 🔗 Read original →

New PET tau‑reading system works with four tracers and detects early Alzheimer’s up to 2.7× more often The HEAD study, published on September 21, evaluated a new five‑class tau‑PET severity scale in 681 participants who received two to four scans with different tau tracers. Three physicians read the images independently, blinded to each other’s scores and to patient data. The regulator‑approved “yes/no” test works only with flortaucipir and is deliberately specific, which causes it to miss early disease. In comparison, the new scale identified 1.9–2.7 times more tau‑positive cases among cognitively normal individuals and 1.2–1.4 times more among those with mild impairment, while results converged at the dementia stage. Development of the method was led by Tariq Pascoal and Suzanne Baker, supported by the US National Institute on Aging with $41 million over five years. Earlier work showed that the MK‑6240 tracer is twice as sensitive as flortaucipir, though each tracer has a distinct binding profile preventing direct comparison. Amyloid PET already uses a common scale; the FDA approved a unified amyloid‑PET scale in August. Tau spreads predictably from deep memory structures to neighboring cortex. A normality threshold was set using scans of fifty amyloid‑negative healthy volunteers across all four tracers. The brain was divided into seven zones ordered by involvement, yielding the five classes: negative, low, moderate, high, and atypical. A free web tool guides readers through the zones and assigns the class; different readers and tracers gave highly concordant ratings. Cases flagged positive only by the new scale showed higher amyloid burden and elevated blood p‑tau217, confirming they represent early pathology rather than false positives. Memory test scores remained stable between negative and low classes but then declined stepwise: 2.5–3.3 points from low to moderate and another 4.4–4.6 points from moderate to high. As Pascoal stated in May, “Tau is the biology most tightly linked to symptoms and future decline. Finding tau earlier and staging it more precisely lets us identify who truly follows an Alzheimer’s trajectory — critical for trial enrollment now and for clinical decisions when new therapies emerge.” 🔗 Read original →

Physical resilience linked to slower kynurenine rise in older adults US and Dutch scientists released a preprint on 21 September describing a 28-year observation of the Rancho Bernardo Study cohort. They measured 22,736 molecules in serum from three visits of 237 participants and related the data to grip‑strength and chair‑rise speed as markers of physical resilience. A statistical model singled out 2,810 molecules associated with resilience from the full panel. Medium‑chain acylcarnitines — carriers of fatty acids into mitochondria — were higher in the low‑resilience group, signalling incomplete fat oxidation, whereas phosphocholines, lipid components of cell membranes, were elevated in those who maintained strength. Among five tryptophan derivatives, only kynurenine showed a different age‑related trajectory between groups: it rose more steeply in participants with low resilience at every visit. In a parallel experiment, activating the OR10J5 receptor in old mice drove mitochondrial growth and restored grip strength to youthful levels. Checking the acylcarnitine signatures against the tissueMASST database showed they are significantly 🔗 Read original →

Organ‑System Divergence in Routine Check‑ups Predicts Mortality Risk Researchers at Zhejiang University built an aging atlas from routine health check‑ups of 189 095 Chinese adults, compiling 94 055 326 measurements of 99 biomarkers across ten organ systems. For 147 492 of these participants, follow‑up mortality data were available, recording 4 599 deaths over the observation period. The analysis showed that the average deviation load across all systems predicted mortality, with each incremental step raising the risk of death by 5,8%. Greater instability between successive check‑ups increased mortality risk by 19% per step, and the effect was almost 1.5‑fold stronger for cardiovascular deaths, though it weakened after stricter statistical correction. Using the same data, the team trained a neural network called DynLifeNet that forecasts seven future health problems—such as kidney dysfunction, high uric acid, hypertension, glucose and cholesterol disturbances, anemia, and liver‑enzyme elevation—one to three years ahead. The model achieved an accuracy of 0,80 (versus 0.5 for random guessing), and adding organ‑system decomposition gave a small but statistically significant boost over baseline measurement history. Validation in an independent cohort of 16 478 participants from another long‑term Chinese study confirmed the relative importance of the organ systems, although the absolute risk scale did not transfer directly. 🔗 Read original →

Hormone therapy started at 46‑56 may cut Alzheimer risk, Stanford and UK Biobank studies show Autopsies, live brain scans, and data from 183 450 women in the UK Biobank converge on the idea that menopausal hormone therapy can lower Alzheimer risk when begun between ages 46 and 56. A Stanford autopsy study published in Neurology, August 2026 found 35% lower odds of Alzheimer pathology among women on estrogen‑only therapy compared with non‑users. A companion UK Biobank analysis in Alzheimer's & Dementia, August 2026 linked therapy to a 10% lower dementia risk and a 16% lower Alzheimer risk overall, with the strongest protection seen in women who started treatment at 46–56 years. The benefit was even greater for those who had undergone oophorectomy (26% lower dementia risk) and for APOEε4 carriers (13% lower risk). These results echo earlier findings that hormone therapy can neutralize the heightened Alzheimer risk conferred by the APOEε4 gene. Twenty years ago the Women's Health Initiative Memory Study (WHIMS) reported that estrogen plus progestin in women aged 65+ nearly doubled dementia risk, prompting US hormone‑therapy use to fall from 27% to under 5%; the WHIMS estrogen‑only arm, however, showed no significant harm. On 12 August 2026 the Stanford team examined post‑mortem brain tissue from 21 462 participants: 258 women had used estrogen‑only therapy and 2701 had not. The odds of Alzheimer signs at autopsy were 35% lower in the therapy group. Senior author Hadi Hosseini noted they evaluated all Alzheimer diagnostics, including the gold‑standard autopsy markers, while co‑author Jennifer Bruno described the effect as modest but significant. The second study, also appearing in Alzheimer's & Dementia, August 2026, followed 183 450 UK Biobank women for 13 years. Therapy users experienced a 10% reduction in dementia risk and a 16% reduction in Alzheimer risk, with greater benefits after oophorectomy (26% lower dementia) and in APOEε4 carriers (13% lower risk). Crucially, protection appeared only when treatment began at 46–56 years, soon after menopause, matching the “critical window” hypothesis that the brain’s responsiveness to estrogen fades after about a decade. The WHIMS discrepancy is explained by its late start age (65+), well past the protective window. A 2024 PET‑blood study suggested estrogen therapy erases excess tau in APOEε4 carriers, offering a possible mechanism. Autopsy, 🔗 Read original →

AI Claude Speeds Ebola Response in DR Congo Outbreak During one of the largest Ebola outbreaks in history, WHO and partners deployed AI Claude to assemble the virus genome, handle reporting, and run simulations. Since May, the Democratic Republic of Congo has been battling a rare and deadly Bundibugyo strain of Ebola for which no vaccine exists. An Anthropic case study from 23 September shows how Claude assists WHO’s African regional office (WHO AFRO), the vaccine coalition CEPI, and Congo’s National Institute of Biomedical Research (INRB). Rural health workers go door‑to‑door in the outbreak zone; their notebooks often remain the only complete record of who fell ill and whom they contacted. Finding patients quickly raises survival chances, but data used to travel slowly: clinics reported cases to districts via WhatsApp, districts fed numbers into PowerPoint slides, and the provincial health ministry compiled them into a daily report — a process that once took an entire day. “I have fought this virus for fifty years. Tools have changed completely, but the rule has not: to beat Ebola you need to know where it is today, not where it was last week,” says Jean‑Jacques Muyembe‑Tamfum, director general of INRB. He was among the first to describe Ebola in Congo’s Yambuku in 1976 and led the development of mAb114, one of the first effective treatments, in 2016. WHO AFRO officer Tendai Musa trained Claude on the needed procedure simply by giving it verbal instructions — no programmers required. Claude now extracts and verifies numbers from district slides against the previous report, preparing a summary in an hour instead of a day. Using the same tool, WHO AFRO analysts model outbreak scenarios: “Before, we could not even start thinking about which outbreak model might be better — we only had time to calculate one,” says Paul Ouma, a colleague of Musa. Now analysts run several scenarios at once, which influences where treatment centers are deployed. There is no approved vaccine for Bundibugyo. Ervebo — the vaccine against the Zaire strain — exists, and early data suggest it may offer partial protection against Bundibugyo. CEPI is testing this hypothesis with Claude: “Claude does not decide which serum samples to analyse to obtain cross‑reactivity data for Ervebo against Bundibugyo — that remains an expert scientific decision,” says Polina Brangel of CEPI. Claude merely consolidates experimental data into a comparable format so experts can reach conclusions faster. In the INRB laboratory, equipment generates millions of genetic fragments in a single run. Previously, assembling the genome required manual terminal commands — a programmer’s task. Claude Science, the science‑focused version of Claude, can, on a plain‑language request, assemble the genome, build an infection tree (who infected whom), and help estimate outbreak scale and spot new variants. The same approach is being transferred by WHO to a chikungunya outbreak — mosquito‑borne fever — in the same region. The next goal is to link historical outbreak databases so new solutions can immediately draw on past experience. According to WHO, by mid‑September over seven thousand cases and more than three and a half thousand deaths had been confirmed in seven eastern provinces of Congo, compared with 4.9 thousand cases and 2.3 thousand deaths by mid‑August when the PARTNERS treatment trial began. This is the second‑largest documented Ebola outbreak in history after the 2014‑2016 West Africa epidemic, which recorded about 28 thousand cases over two years; the current outbreak has reached a quarter of that figure in just a few months amid ongoing armed conflict. 🔗 Read original →

Cell death mechanisms expand beyond apoptosis and necrosis The number of known ways a cell can die has risen from the two classic pathways — apoptosis and necrosis — to more than twenty distinct mechanisms, most of which were uncovered in the last 27 years. Each route serves a specific role: some trigger an alarm for the immune system, others proceed silently, and occasionally a cell that has begun to die can recover. New pathways are reported almost every year; at least four were described in 2025 and another emerged this spring. For a long time, cell death was divided into disorderly necrosis — where a cell is crushed, cut, or poisoned and falls apart without order — and controlled apoptosis, in which the cell dismantles itself into membrane blebs that the immune system clears. Advances such as gene editing, high‑resolution microscopy, and multi‑molecule analysis have exposed mechanisms that do not fit the old dichotomy. A Nature article dated September 22 gathered them together: since 1999 scientists have outlined about twenty new mechanisms, and the list continues to grow. Cell biologist Daolín Tan discovered alkaliptosis while screening compounds to kill cancer cells; he observed a sudden spike in intracellular pH that made cells overly alkaline and lethal. He named the process alkaliptosis. The list includes pyroptosis and necroptosis — explosive, pro‑inflammatory routes; ferroptosis — driven by iron accumulation that destroys membranes; cuproptosis — triggered by excess copper; and sodium‑overload death, first described in 2025. This spring researchers found another pathway in flatworms: exploding “ruptoblast” cells that spray toxin on bacteria. The mode of death determines what happens to the tissue afterward. Biophysicist Ana García‑Saéz explains that maintaining many parallel routes is essential: if a pathogen blocks one death scenario, the infected cell needs a backup. “It’s like a race 🔗 Read original →

AI 🔗 Read original →

Claude AI agents discover novel ART system in phage genomes resembling CRISPR Anthropic announced on September 23 that it set roughly 950 Claude instances the task of searching genomic databases for new reverse transcriptases. After 21 hours one agent noticed a DNA pattern next to an unusual enzyme that resembled CRISPR and reported the finding to scientists. The lab confirmed the DNA segment is transcribed into RNA, though the exact function remains unknown. Agents were then assigned a routine task to gather reverse transcriptases and locate neighboring genes. One agent abandoned its initial neighbor hypothesis and decided to examine the DNA upstream for a short RNA, similar to retrons. Reading the raw sequence, the agent logged astonishment at seeing a tandem repeat array akin to CRISPR, counted the repeats, measured their spacing, and consulted the literature. Researchers had previously observed the locus; another scientist had described the transcriptase and guessed a nearby small RNA but missed the repeat array and partner gene. The authors note that genomic data now exceed what scientists can monitor. Re‑examining the same region, the agent 🔗 Read original →

AI model EAGLE detects esophageal cancer on routine non‑contrast CT scans On 22 September Nature Medicine, September 2022 published an article describing the EAGLE model, which identifies esophageal cancer and precancerous lesions on standard contrast‑free CT. Validation in eight centers across three countries involved 11,466 patients and yielded 89.5% sensitivity with only a 1.5% false‑positive rate. The system is already in use at three Chinese hospitals, where it flagged cancers missed by conventional diagnostics. Globally, esophageal cancer accounts for about 511,000 new cases and 445,000 deaths each year, with five‑year survival of 36.9 % in China and 18.5 % in the United States. No widespread screening exists because endoscopy is invasive and limited to high‑risk regions in China. Many individuals undergo chest CT without contrast for other indications; such scans capture the entire esophagus and represent roughly 40 % of all CT examinations worldwide. Early mucosal changes are subtle, and radiologists often overlook them—most patients who later died of esophageal cancer had no suspicious findings on prior CTs, leaving the signal buried in archives. EAGLE was constructed by Alibaba DAMO Academy and Chinese oncology centers in two stages: one neural network isolates the esophagus from the scan, a second searches that region for malignant or precancerous alterations. Training used voxel‑wise tumor outlines from contrast CT, transferred to plain images via image registration, allowing the model to discern faint early signals. In testing, EAGLE surpassed each of 17 radiologists; with its assistance, clinicians detected cancer in 85.7 % of patients versus 71.9 % before. Sensitivity for early‑stage tumors was lower (60‑66 %) than for established lesions (89‑90 %). The model also maintains 88.4 % sensitivity on low‑dose lung‑cancer screening CTs, enabling simultaneous esophageal cancer detection without extra procedures. Clinical deployment showed promise: an initial wave in three Chinese hospitals (21,000 patients) uncovered three cancers missed by routine work‑up, one 21 months earlier than the official diagnosis. 🔗 Read original →