FAHR-Face stands for Foundation Artificial Intelligence for Health Recognition. Its design follows the foundation-model approach: a large system first learns a general representation from a broad collection of face images and is then adapted to particular health-related outcomes. Rather than training an independent algorithm from scratch for each task, the researchers used the shared model to develop both an age estimator and a survival predictor. This structure also allowed them to compare the two outputs directly and ask whether biological-age information and survival information were redundant. The study was computational rather than a clinical trial: it developed and validated prediction models but did not test whether physicians using the scores made better treatment decisions or whether the tools changed patient outcomes.

FAHR-FaceAge was reported to estimate biological age more accurately than previous age-estimation algorithms. The researchers evaluated the model with photographs subjected to common variations encountered in real-world imaging, including differences in pose, lighting, facial expression and makeup. Results remained consistent across most of these conditions, indicating that the age estimates were not restricted to a single photographic setup. The model was also stable across most plastic-surgery procedures included in the analysis. A small shift was observed following rhinoplasty, while the other tested procedures did not produce comparably large changes in the reported estimates. These findings concern the model’s stability under the specific variations and procedures tested; they do not establish that its performance is unaffected by every possible image or medical condition.

FAHR-FaceSurvival takes a different approach from estimating an age-related characteristic and then testing its relationship with survival. It was developed to learn cancer-survival outcomes directly from facial photographs. The age and survival predictions were only weakly correlated, suggesting that the two tools did not simply reproduce the same output. When the measures were analyzed together, each remained independently associated with patient survival. Combining them also improved the fit of the survival model in the external validation cohort of cancer patients. This added model fit is evidence that the two facial-derived measures may contain complementary prognostic information, but the study did not establish a clinical decision threshold, treatment-selection rule or safe way for either score to be used on its own.

Visual analysis of the models showed that they assigned the greatest influence to different parts of a face. FAHR-FaceAge focused on the nasolabial folds, the forehead immediately above the eye sockets and the temples. FAHR-FaceSurvival instead emphasized the area below the eyes and the region around the nasal bridge. The researchers used these maps to show which facial regions most influenced the respective estimates and predictions. The differing regions accompanied the weakly correlated outputs, but the study did not identify a biological mechanism linking particular facial areas to either biological aging or cancer survival. Accordingly, the maps describe the models’ use of visual information rather than proving that the highlighted areas are direct measures of a specific physiological process.

The analysis also identified associations between lifestyle exposures and the biological-age estimates. Heavy alcohol use, drug use and smoking were each associated with looking biologically older. Smoking showed the clearest pattern: current smokers appeared older than former smokers, and former smokers appeared older than people who had never smoked. These were associations observed in the study data rather than evidence that smoking or other behaviors caused the estimated differences or that changing those behaviors would reverse the model’s estimate. The results also do not show that the age tool was intended to diagnose lifestyle exposures. Instead, they indicate that facial-age estimates varied in relation to reported behaviors, alongside variation associated with chronological age.

In examining prognosis, the investigators asked whether the direct survival score and the difference between FAHR-FaceAge and chronological age contributed information beyond one another. Both measures were independently associated with survival when analyzed in the same model, and their combination improved model fit in the external cancer cohort. This supports the interpretation that the age and survival tools can represent related but separable information within the study. It does not mean that a facial photograph alone reveals the cause of a patient’s prognosis, nor does it establish that biological age functions as a mechanism leading to shorter survival. The reported result concerns statistical association, predictive performance and incremental model fit in the populations included in the development and validation work.

Raymond Mak, a radiation oncologist at Mass General Brigham Cancer Institute and faculty member in the institution’s Artificial Intelligence in Medicine program, said the broader goal is to make health assessment from face photographs more quantitative, consistent and scalable. He described this objective in the context of estimating physical reserve and frailty before physicians consider intensive treatments such as radiation, surgery or systemic therapy. The current study, however, did not compare AI-guided treatment selection with standard care or measure safety, treatment response or downstream clinical benefit. Mak and co-senior author Hugo Aerts presented the tools as potential support for personalized treatment decisions rather than replacements for physician judgment. Their findings therefore support further investigation of the scores rather than a demonstrated clinical practice change.

Aerts said FAHR-FaceSurvival advances facial health-recognition research by learning survival risk directly from photographs of patients with cancer. The team described the result as the first demonstration, to its knowledge, of an AI model directly predicting cancer survival from a simple photograph. That direct-learning approach distinguishes the survival tool from an age-estimation model used as an indirect feature. The evidence presented nevertheless remains within the boundaries of a model-development and validation study. No specific biological pathway, patient subgroup in which the tool was safe to deploy, or standard against which AI-assisted care was compared was established in the release.

The researchers are extending the work through the FaceAge Research Study, an online study approved by an institutional review board and open to the public. Participants can upload a facial photograph, receive a FaceAge research readout and contribute data intended to help develop and validate AI-based measures of facial aging and health in larger and more diverse populations. Additional evaluation will be needed to determine how the models perform across broader demographic and clinical settings, beyond the robustness already observed for pose, lighting, expression, makeup and most tested plastic-surgery procedures. The study’s supported conclusion is that FAHR-Face can derive biological-age estimates and direct cancer-survival predictions from facial photographs, with the combined measures adding prognostic value to the survival model in external validation. It is not yet evidence that using those predictions improves treatment choices or clinical outcomes.

Subject of Research: Development and validation of facial-photograph AI models for biological-age estimation and cancer-survival prediction

Article Title: AI facial health recognition tools help physicians assess age and survival risk to guide personalized treatment

Article References: AI facial health recognition tools help physicians assess age and survival risk to guide personalized treatment. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: facial artificial intelligence, biological age, cancer survival, foundation models, computer vision, medical imaging, prognostic modeling, external validation, digital health, patient photographs, facial, health

The review, led by Guangjun Wang and colleagues at Anqing Normal University, Shanghai University of Traditional Chinese Medicine and Xiangyang Central Hospital, synthesizes the entire pipeline of AI-driven biological age (BA) prediction: which biomarkers feed the models, how engineers transform raw omics and imaging data into learnable features, which algorithms dominate, how systematic bias is corrected, and how performance is judged. Its central message is technical but consequential: aging is not a single number. It is an asynchronous process in which the brain, blood vessels, kidneys and immune system each age at their own rate, and only modern machine learning — particularly deep learning and multimodal fusion — has the capacity to model such nonlinear, system-specific trajectories.

The biological raw material for these models spans a continuum. At the molecular tier sit DNA methylation clocks — Horvath, Hannum, PhenoAge and GrimAge — which read chemical tags on the genome and remain the most accurate single-modality estimators of biological aging, showing robust associations with mortality, disease risk and healthspan. Telomere length reflects replicative senescence; transcriptomic and metabolomic signatures capture age-related shifts in gene regulation and metabolism, with machine learning models trained on multi-omic data reaching prediction errors of roughly four to five years. At the system tier, blood-based panels — inflammatory markers such as C-reactive protein and interleukins capturing the phenomenon of inflammaging, alongside metabolic, hepatic, renal and hematologic indices — are transformed into BA estimates by statistical methods like the Klemera–Doubal method and by deep learning blood clocks, typically within about five years of error.

Imaging adds an organ-specific dimension that molecular clocks cannot provide. Brain MRI models quantify cortical thinning, gray and white matter degradation and network disorganization, with deep learning architectures achieving mean absolute errors as low as two to three years and strong links to cognitive decline. Dental and skeletal X-rays quantify mineralized tissue aging; facial imaging captures cutaneous and craniofacial aging shaped by lifestyle; and retinal photographs offer a non-invasive window into microvascular and neural aging, scalable for population screening. These modalities are ideal for AI precisely because they are high-dimensional and plentiful — the kind of data convolutional neural networks and transformers were built to digest.

Feature engineering, the review stresses, is not a mere preprocessing chore but a design decision that determines whether a model can capture asynchronous aging at all. High-dimensional methylation loci and metabolomic profiles require normalization and batch-effect mitigation; principal component analysis and autoencoders distill compact latent factors from noisy omics; classical screens such as correlation with chronological age, Cox proportional hazards regression and variance inflation factor diagnostics still underpin transparent, clinically intuitive biomarker panels. The AI shift comes with representation learning: variational autoencoders learn noise-robust embeddings, graph neural networks model biomarker dependencies such as gene co-expression networks, and attention-based transformers learn feature importance dynamically — architectures well suited to encoding organ-specific representations and fusing them into a unified estimate that respects differential aging rates across subsystems.

On the modeling side, the field has progressed from weighted regression schemes like the Klemera–Doubal method — interpretable but constrained by linear assumptions — through random forests, support vector machines and elastic net regression, to deep architectures that learn aging representations end-to-end. The milestones are striking. DeepMAge, a neural network epigenetic clock trained on more than 4,000 blood methylomes, achieved a mean absolute error of 2.77 years, outperforming regression baselines. In neuroimaging, a global–local transformer framework reduced brain age error to 2.7 years. Multimodal transformers now integrate methylation, transcriptomic and imaging signals; contrastive learning disentangles aging signals from confounding noise; and multitask learning jointly predicts BA alongside sex, disease risk or mortality to improve regularization and transferability.

Yet raw accuracy is not enough, and the review devotes unusual care to model correction — the post-prediction step that separates a publishable number from a clinically trustworthy one. Uncorrected BA models systematically overestimate age in younger individuals and underestimate it in older ones, the notorious age–delta correlation or regression-to-the-mean problem. Skewed loss functions penalize these asymmetric errors during training, as in the AccelerAge framework, where gradient magnitudes adapt to a subject’s age and error direction, driving residual–age correlations toward zero on Cam-CAN and ABIDE neuroimaging datasets. AgeML performs transparent linear recalibration of predicted against true age; z-score correction standardizes outputs to population-level means and variances; and risk-difference methods define BA relative to expected age-dependent physiological profiles, as in vascular age frameworks built from radial artery pulse waves.

Evaluation is equally two-sided. The data perspective leans on mean absolute error, RMSE and the coefficient of determination — deep learning models in neuroimaging and multimodal settings now routinely achieve RMSE below three to four years. But high numerical performance does not guarantee biological validity. The knowledge perspective asks whether BA deltas — the gap between predicted and chronological age — track real health differences: cross-sectional analyses link accelerated aging to disease, frailty and socioeconomic conditions, while longitudinal cohort studies, the stronger test, show that BA acceleration predicts cognitive impairment, multimorbidity and mortality. Electronic health records extend this validation to population scale, with AI-predicted ECG age gaps correlating with mortality and cardiovascular risk after bias adjustment.

The challenges, however, are formidable, and the review is candid about them. Biomarker panels vary wildly across studies, limiting reproducibility; aggregating heterogeneous signals into a single BA output can obscure organ-level aging patterns. Most supervised models train against chronological age, implicitly assuming that deviations from CA reflect true aging — an assumption that can bias models toward chronological patterns rather than aging mechanisms and invites overfitting. Deep models function as black boxes, complicating clinical acceptance where interpretability and causal insight are prerequisites. Cross-sectional studies dominate the literature but cannot capture temporal trajectories or asynchronous dynamics. And most current models ignore genotype altogether: variants in APOE and FOXO3 demonstrably modulate aging trajectories — the APOE4 allele is linked to faster biological aging while FOXO3 polymorphisms show reproducible longevity associations — yet polygenic risk scores and longitudinal clinical history are rarely built into model architectures, risking the conflation of disease burden with intrinsic aging.

The path forward, the authors argue, requires standardized preprocessing protocols, robust validation across demographically diverse populations, interpretable and bias-aware AI, and above all the integration of multi-omics and longitudinal datasets into frameworks that model aging as the multidimensional, asynchronous process it is. If the field succeeds, biological age could graduate from a descriptive curiosity — the number behind longevity headlines — into a clinically actionable metric: a system-level dashboard of how fast each of your organs is aging, and a target for interventions that might slow it. The machines are already reading the clock. The task now is to make sure they are reading it fairly, transparently and in a way that changes what happens in the clinic.

Subject of Research: AI biological age prediction review

Article Title: Artificial intelligence approaches in biological age prediction: current status and challenges

Article References: Wang, G., Ding, P., Li, Z., Tang, Q., Tu, L., Jiang, T., Zhang, L., Su, B., Xu, J., & An, H. (2026). Artificial intelligence approaches in biological age prediction: current status and challenges. British Journal of Biomedical Science, 83, Article 16141. https://doi.org/10.3389/bjbs.2026.16141

Image Credits: AI Generated

DOI: 10.3389/bjbs.2026.16141

Keywords: aging, artificial intelligence, biological age, epigenetic clocks, deep learning, precision medicine, Artificial, intelligence, approaches, biological, prediction, current

The research topic, titled Biomarkers and Aging: New Insights for Precision Medicine, gathers contributions that trace a deliberate arc, from the conceptualization of biological age to its molecular measurement, computational interpretation and eventual clinical relevance. Its editors, Oluwafemi Gabriel Oluwole of Olabisi Onabanjo University and the University of Cape Town and Paulo Adriano Schwingel of the University of Pernambuco in Brazil, argue that aging is a dynamic, multidimensional phenomenon shaped by genetic and epigenetic regulation, inflammation, immune and stress responses, DNA repair and individual exposures. No single biomarker or score, they contend, can capture the whole picture.

The most influential tools in this field have come from an unexpected source: chemical tags on DNA. DNA methylation, the attachment of methyl groups to cytosine bases across the genome, changes in predictable patterns as cells age, and these patterns can be read out as so-called epigenetic clocks. A scoping review by Ziesel and colleagues examined 435 studies in this space, 167 of which used methylation-based approaches, charting the progression from simple age-associated CpG patterns to increasingly sophisticated clock models. A companion review from the same group extended the analysis to 309 non-methylation-based biomarkers, revealing a bewildering diversity of candidate markers and metrics.

Read together, the two reviews signal a decisive shift in how the field thinks about measurement. The era of hunting for the single best biomarker of aging is giving way to a search for combinations that capture the dimensions of aging relevant to a specific individual and a specific clinical outcome. The editors illustrate why this matters with a striking asymmetry: a person may show accelerated epigenetic aging alongside fully preserved cardiovascular function, or conversely display metabolic and immune aging despite a modest global biological-age estimate. A single number, in other words, can inform while simultaneously concealing real and clinically meaningful heterogeneity.

Biological age would remain an academic curiosity if it did not connect to outcomes that patients and physicians care about. Chen and colleagues addressed this directly, drawing on data from 349,456 participants in the UK Biobank. Over roughly 14 years of follow-up, they found that phenotypic age acceleration partially mediated the association between impaired lung function and both mortality and cardiovascular disease, with risk highest in participants showing both exposures. The finding suggests that biological age can capture accumulated vulnerability that chronological age or a single organ measurement misses, acting as an integrative indicator of interacting processes rather than a direct cause of disease.

The UK Biobank analysis also carries an important caveat that the editors emphasize: biological-age measures are best understood as complements to established risk factors, not replacements for them. A clinician who abandons conventional cardiovascular risk assessment in favor of an epigenetic score would be discarding decades of validated evidence. The more defensible vision is a layered one, in which biological-age estimates sit alongside traditional measures to refine prediction, particularly in the gray zone where conventional risk calculators disagree with clinical intuition.

Proteomics adds a functional dimension to this layered picture. While epigenetic marks are regulatory signals, proteins are the functional effectors that actually execute biology, so proteomic biomarkers offer a window into downstream biological activity that DNA methylation cannot provide. Luo and colleagues reviewed proteomic biomarkers spanning the immune, metabolic, endocrine, cardiovascular, musculoskeletal and nervous systems, together with proteomic aging clocks linked to frailty, multimorbidity, organ-specific biological age and mortality. Their synthesis reinforces the emerging consensus that a multidimensional aging profile, rather than a single biological-age estimate, is the realistic goal.

Physiological signals point in the same direction. In their own contribution, Oluwole and Schwingel found that reduced heart-rate-variability indices, a marker of autonomic and parasympathetic modulation, were associated with cognitive decline in community-dwelling older adults. Heart-rate variability can be measured with inexpensive, non-invasive equipment, which makes it an attractive functional complement to expensive molecular assays. The finding illustrates how system-specific functional biomarkers extend, rather than replace, molecular measures of biological age, adding real-time information about how the aging nervous system and cardiovascular system are actually performing.

Moving from measurement to precision medicine also requires linking molecular information to candidate interventions, and here computational methods are proving decisive. The editors describe how transcriptomic analysis of human Achilles tendon tissue, combined with sequence analysis, molecular docking and protein-protein interaction analysis, was used to investigate a small peptide derived from methyl jasmonate esterase and its interaction with human complement factor H. In related work, network pharmacology and molecular docking applied to Sorghum bicolor 3-deoxyanthocyanidins identified inflammation- and redox-related pathways, including NF-κB, Keap1-Nrf2 and PPAR-γ, with experimental support for modulating inflammaging-related processes. Neither example constitutes an established anti-aging therapy, but both illustrate an emerging pathway in which biomarkers and mechanisms help prioritize hypotheses for further, experimentally validated investigation.

Two open questions will determine whether this translational framework delivers on its promise. The first concerns population diversity: whether biomarkers and computational clocks developed in specific populations generalize across genetic backgrounds, environments and socioeconomic circumstances remains unresolved, and precision medicine cannot be precise if its underlying biomarkers under-represent the populations in which they will be applied. The second concerns actionability: a biomarker that predicts disease is scientifically valuable, but its clinical worth depends on whether it identifies modifiable processes and supports interventions that improve outcomes. That will require longitudinal studies tracking whether biological-age measures respond to lifestyle or pharmacological intervention, and whether those changes translate into meaningful gains in healthspan. The editors’ conclusion is measured but ambitious: the field is moving beyond measuring biological age toward understanding biological vulnerability, and the next stage will depend on multi-omics measurement, computational modelling and diverse longitudinal cohorts, with models kept reproducible, interpretable and clinically validated.

Subject of Research: Biomarkers of biological aging and their application in precision medicine

Article Title: Editorial: Biomarkers and aging: new insights for precision medicine

Article References: Oluwole, O. G., & Schwingel, P. A. (2026). Editorial: Biomarkers and aging: new insights for precision medicine. Frontiers in Aging, 7, Article 1987600. https://doi.org/10.3389/fragi.2026.1987600

Image Credits: AI Generated

DOI: 10.3389/fragi.2026.1987600

Keywords: biological age, biomarkers, epigenetic clocks, DNA methylation, proteomics, heart-rate variability, precision medicine, aging, inflammaging, UK Biobank, frailty, computational biology

The work was led by Alex Zhavoronkov, a member of the Aging editorial board who is affiliated with Insilico Medicine and the Buck Institute for Research on Aging. Its starting observation is blunt: established frameworks such as the Hallmarks of Aging and Mikhail Blagosklonny’s hyperfunction theory explain mechanisms of decline, but they do not prescribe how, when or whether those mechanisms should be targeted therapeutically. “Existing aging theories describe what changes with age but do not prescribe how to intervene,” the authors write. Control theory, the mathematical discipline concerned with directing dynamic systems toward desired states, offers precisely that prescriptive language.

At the heart of the framework lies the concept of a biological state space, a multidimensional representation of an organism’s condition spanning inflammation, mitochondrial function, epigenetic stability, immune competence and regenerative capacity. Rather than attempting to rewind every molecular feature to a youthful value, the framework prioritizes maintaining or recovering acceptable physiological function. The authors define a functional viability region, the set of biological states in which essential functions remain within acceptable limits, surrounded by a wider recovery region from which safe interventions could potentially restore function. As aging progresses, this recoverable territory may shrink, making functional restoration progressively harder and eventually impossible.

From this geometry follows a striking redefinition of two familiar concepts. Aging, in this framework, is a progressive loss of safe controllability, and biological age is interpreted not as a molecular signature but as the minimum safe cost required to restore or maintain function. This distinguishes the approach from conventional biological aging clocks, which generally estimate chronological age or health-related risk from molecular measurements. Here, the measure of age is practical: how difficult would it be, with the interventions actually available, to achieve meaningful functional improvement without pushing the system into a dangerous state?

Drugs enter the mathematics as vector fields, meaning mathematical descriptions of how an intervention changes a biological system’s trajectory through state space. Because a vector field depends on the current state, the same treatment can push one patient toward recovery and another toward harm. Treatment effects therefore depend on biological state, dose, duration and tissue conditions rather than being universally beneficial. The authors argue that this formulation could help explain why individuals of similar chronological age may respond so differently to the same intervention, a persistent puzzle in geriatric medicine and in trials of aging-related therapeutics.

The framework also delivers a formal account of why treatment order matters. Using mathematical objects known as Lie brackets, which capture how the composition of two influences depends on the sequence in which they are applied, the authors propose that administering two interventions in different sequences can produce different biological effects even when the same therapies are used. To demonstrate the principle, they built a five-variable computational model of an aged mouse liver incorporating senescent-cell burden, oxidative damage, epigenetic integrity, regenerative capacity and fibrosis, and compared three hypothetical strategies: senolytic treatment followed by partial cellular reprogramming, the reverse sequence, and simultaneous administration.

In the illustrative 56-day simulation, senolytic treatment before reprogramming produced the most favorable modeled outcome, with the proposed explanation that clearing senescent cells first creates conditions that improve subsequent reprogramming. The authors are careful about what this proves. These findings came from mathematical simulations rather than animal experiments, and the predicted benefits of treatment sequencing require biological validation before any clinical conclusions can be drawn. The value of the exercise lies in showing that sequencing effects, often treated as anecdotal, can be made explicit, testable predictions within a rigorous formal structure.

Safety occupies the center of the framework rather than its periphery. “Safety is not a secondary penalty appended after efficacy; it is the central constraint that defines the admissible control set,” the authors write. They distinguish between reversible interventions, such as many small-molecule treatments, and potentially durable approaches including gene editing, viral gene delivery and certain cell therapies. Under the framework, reversible interventions are prioritized when uncertainty is high, while longer-lasting therapies are considered only when their expected benefits and safety are sufficiently supported. The approach also acknowledges hard biological boundaries: extensive fibrosis or irreversible cell loss may mark points beyond which available treatments can no longer restore normal function, underscoring the potential importance of intervention timing and of preventive strategies over late-stage rescue attempts.

To explore clinical relevance, the researchers outlined three translational case studies covering immune aging, age-related muscle loss and ovarian aging, illustrating how intervention selection could account for tissue-specific changes, functional goals and safety constraints. A recurring theme is that molecular proxies can mislead: increasing muscle mass, for example, would not constitute successful treatment if strength, mobility and metabolic function remained impaired. The paper also applies an exploratory scoring framework to aging-related interventions across five biological stages, drawing on aging research databases, clinical trials and therapeutic-target resources, and evaluating treatments by biological coverage, safety, reversibility, durability and potential functional benefits. In this exercise, caloric restriction and aerobic exercise ranked highest, followed by the SGLT2 inhibitors dapagliflozin and empagliflozin, with GLP-1 receptor agonists including liraglutide and semaglutide also scoring relatively high. The authors stress that these scores depend on the framework’s assumptions and weighting criteria and do not demonstrate that any of these drugs extend human lifespan or should be prescribed specifically to treat aging.

Looking forward, the paper sketches a proposed computational system that would integrate molecular measurements, clinical data, mathematical modeling and repeated monitoring to estimate biological states and adjust interventions over time, an approach that could eventually support individualized therapeutic decisions, although a fully trained and experimentally validated implementation does not yet exist. The authors also present 20 falsifiable predictions for future investigation, including whether treatment sequencing influences functional restoration, whether biological states predict individual responses, and whether safety-constrained strategies outperform approaches based mainly on aging biomarkers or single molecular targets. The limitations are acknowledged plainly: biological states cannot be measured completely, the models rest on assumptions requiring experimental validation, the rankings depend on chosen scoring criteria, and the framework’s advantages have not been demonstrated in prospective studies. What the paper offers, ultimately, is a shift in attention for aging drug discovery, away from cataloguing age-related changes and toward determining how function might be restored safely, delivered as a set of testable principles that independent laboratories can now begin to challenge.

Subject of Research: A control theory framework for aging biology and safer gerotherapeutic drug discovery

Article Title: New control theory framework proposes a safer approach to aging drug discovery

Article References: New control theory framework proposes a safer approach to aging drug discovery. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: aging, control theory, gerotherapeutics, drug discovery, biological age, senolytics, cellular reprogramming, state space, vector fields, safety, caloric restriction, SGLT2 inhibitors

Conventional comorbidity indices such as the Charlson Comorbidity Index, the Elixhauser score, the Functional Comorbidity Index, the Hospital Frailty Risk Score, and the Multimorbidity Weighted Index are widely used in hospitals and research settings. These instruments typically depend on fixed disease lists and fixed weights derived from specific populations. That design choice creates a persistent problem: when such scores are transported to a different country, a different healthcare system, or a different age structure, their calibration and predictive performance can deteriorate. A weighting scheme tuned to older European patients may misjudge risk in younger Asian cohorts, limiting genuinely transportable risk stratification across heterogeneous populations.

Large language models offer a fundamentally different computational approach. Rather than counting diagnoses, an LLM can integrate heterogeneous health information, including medical history, laboratory measurements, lifestyle factors, functional status, and social context, and approximate the cross-system reasoning that experienced clinicians perform intuitively. The research team, led by Jinghao Liang of Guangzhou Medical University, set out to test whether this capability could be converted into a stable, interpretable multimorbidity phenotype score. However, the authors were careful to acknowledge the caveats that accompany medical artificial intelligence. Medical LLMs face concerns regarding reliability, bias, and cross-population generalizability, which makes large-scale external validation essential before any clinical deployment.

The resulting framework, called the LLM-based multimorbidity score, or LLM-MMS, was developed and validated for all-cause mortality risk stratification using nine international longitudinal cohorts: the UK Biobank, NHANES, SHARE, HRS, CHARLS, CLHLS, ELSA, KLoSA, and MHAS, together comprising 770,439 participants. The pipeline begins with harmonization. Demographic characteristics, lifestyle factors, physical examinations, biomarkers, and medical histories from all cohorts were mapped through a unified data dictionary and converted into standardized English natural-language clinical reports, with explicit hallucination control during the textification step. This conversion of structured data into narrative form is what allows a language model to read a patient profile the way a physician would.

At the core of the system sits DeepSeek-V3, used as the reasoning engine with a fixed prompt template and chain-of-thought prompting. Notably, the model was not fine-tuned on any clinical data. Through prompt-based learning alone, it generated seventeen health phenotypes for each participant, including an overall biological age, multiple organ-specific biological ages, a frailty age, and multidimensional health gradings covering psychological health, diet, behaviors, medical history, income, and family history. These generated phenotypes were then entered into standard Cox proportional hazards models for mortality prediction. The score was developed in the UK Biobank and externally validated in the other eight cohorts, with head-to-head comparisons against the Charlson, Elixhauser, FCI, HFRS, and MWI indices using the C-index, calibration plots, Brier scores, and decision-curve analysis.

The performance gap was striking. In the UK Biobank, LLM-MMS reached a C-index of 0.81, substantially higher than the best conventional comorbidity index at 0.61. Across the eight external validation cohorts, C-indices ranged from 0.67 to 0.86, exceeding even recalibrated conventional indices by 0.12 to 0.21. Calibration remained stable, with observed-to-expected ratios generally falling between 0.94 and 1.04, meaning the model’s predicted risks closely matched actual mortality rates. Stratified analyses showed relatively consistent discrimination across sex and age groups, with particularly clear advantages among adults younger than sixty years, a population in which traditional comorbidity scores often exhibit floor effects because younger patients accumulate few listed diagnoses even when their overall health trajectory is concerning.

Two further tests addressed the most common objections to LLM-based medicine. First, LLM-MMS outperformed survival models trained directly on the same raw structured variables in the UK Biobank, suggesting that its advantage was not simply a consequence of richer input information but of the model’s ability to transform multidimensional signals into a more holistic phenotype representation without fixed weighting schemes. Second, portability and reproducibility were tested by swapping in alternative LLM backbones and repeating inference runs; performance varied only modestly across different base models and remained highly reproducible, indicating that the framework does not depend on one proprietary engine or on stochastic quirks of individual inference passes.

The biological plausibility of the score was examined through proteomic profiling, model-generated reasoning summaries, and input-masking analyses. High-risk individuals carried molecular signatures related to inflammation, tissue injury, and metabolic stress, precisely the processes one would expect to underlie elevated mortality. Input-masking experiments, in which individual variables were systematically hidden from the model, identified smoking, inflammatory markers, systolic blood pressure, body mass index, and HbA1c as the major contributors to the final risk score. These contributors align with well-established epidemiological risk factors, supporting the interpretation that the language model had learned clinically meaningful patterns rather than exploiting spurious correlations hidden in the harmonized data.

The significance of the work lies in moving beyond the conventional paradigm of disease counting plus fixed weights. By integrating cross-organ and multidimensional health information into an interpretable phenotype, LLM-MMS maintained more stable discrimination and calibration across countries, age groups, and sexes without cohort-specific retraining, suggesting that LLM-derived integrated health representations may offer genuinely improved transportability across heterogeneous populations. The authors are candid about the limitations. LLM-MMS functions more like a broad prognostic health phenotype than a purely diagnosis-based multimorbidity score, and the inclusion of socioeconomic and other non-diagnostic domains raises legitimate questions about fairness and appropriate use. The framework also remains anchored to its UK Biobank development cohort, to English-language textification, and to potential biases inherent in large language models themselves.

Looking forward, the team outlines a roadmap toward safe clinical integration: prospective clinical validation, testing in non-English environments, continuous calibration, fairness audits, and the establishment of model version management and post-deployment monitoring mechanisms. The goal, as Jinghao Liang describes it, is to promote the safe use of such systems for human-machine collaborative risk stratification, in which the AI-generated phenotype serves as a decision-support layer alongside clinician judgment rather than a replacement for it. The paper, Bio-Inspired Artificial Intelligence Multimorbidity Score for Human-Machine Clinical Risk Stratification, was published in Cyborg and Bionic Systems on September 30, 2026, and the work was supported by the Natural Science Foundation of Guangdong Province, the National Natural Science Foundation of China, and the R&D Program of Guangzhou National Laboratory.

Subject of Research: A large language model-based multimorbidity score for all-cause mortality risk stratification across international cohorts

Article Title: Bio-inspired artificial intelligence multimorbidity score for human–machine clinical risk stratification

Article References: Bio-inspired artificial intelligence multimorbidity score for human–machine clinical risk stratification. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: large language models, multimorbidity, risk stratification, all-cause mortality, UK Biobank, biological age, comorbidity indices, DeepSeek-V3, clinical decision support, artificial intelligence, proteomics, population aging

The trial in question, reported by Alex Zhavoronkov and colleagues in Nature Biotechnology, marks the first time that six distinct proteomic aging clocks have been deployed together as readouts in a clinical study of an anti-aging intervention. That detail matters because clinical trials are unforgiving environments for biomarkers. A clock that looks impressive when trained on cross-sectional blood samples from tens of thousands of people may behave very differently when asked to detect a modest, treatment-induced change within individuals over months. Robertson and Chandra’s verdict on the exercise is pointed: the deployment revealed more about the clocks themselves than about the drug being tested. In other words, the trial functioned less as a decisive test of a therapy and more as a stress test of the measurement instruments.

To understand why this matters, it helps to consider how proteomic clocks are built. Researchers start with large cohorts in which plasma protein levels have been measured, typically using affinity-based platforms that quantify thousands of proteins simultaneously. Statistical models, often penalized regressions or machine-learning approaches, are then trained to predict chronological age from those protein patterns. The difference between the predicted age and a person’s true age, the so-called age gap, is interpreted as a measure of biological aging. Large-scale efforts published in recent years, including analyses of UK Biobank data by Argentieri and colleagues in Nature Medicine in 2024 and independent work by Xu and colleagues in 2025, showed that such proteomic clocks predict mortality and incident disease with genuine skill, in some cases capturing organ-specific deterioration that epigenetic measures miss.

But prediction in a population is not the same as sensitivity to change in an individual. A clock can be highly accurate at ranking people by age and still be nearly useless for detecting whether a six-month intervention slowed anyone’s aging. Robertson and Chandra frame the trial through this distinction, which echoes a long-standing concern in clinical medicine formalized by Fleming and DeMets in a 1996 paper in Annals of Internal Medicine: a surrogate endpoint must capture the effect of treatment on the outcome of interest, not merely correlate with the outcome in observational data. Correlation across individuals does not guarantee that a biomarker will respond, or respond in the right direction, when a drug perturbs the underlying biology. This is the central technical challenge now facing proteomic geroscience.

The trial also exposed subtler problems that arise when multiple clocks are run on the same samples. Different proteomic clocks are trained on different platforms, different cohorts and different sets of proteins, and they can disagree. Two clocks might both predict chronological age well, yet one might show a treatment effect while another shows none, or even a shift in the opposite direction. Without a validated gold standard, it is difficult to know which clock, if any, is tracking true biological aging rather than platform noise, inflammation from an intercurrent illness, or compositional changes in blood cell populations. Robertson and Chandra emphasize that the six-clock comparison turned this ambiguity into data, allowing the field to see where the clocks converged and where they diverged when confronted with real longitudinal samples from a controlled study.

There is also a deeper statistical issue lurking beneath the age-gap framework. Proteomic clocks are trained to predict chronological age, so by construction the model tends to pull predictions toward the population mean. If a treatment genuinely shifts a person’s protein profile toward a younger state, the clock should register a negative age gap. But the same shift can be produced by regression to the mean, by changes in the overall variance of the proteome, or by the treatment altering the levels of a handful of highly weighted proteins without broadly rejuvenating the system. Dabrowski and colleagues explored related questions in Nature Aging in 2024, examining how clock construction choices shape what the resulting scores actually measure. The trial setting, with its randomized structure and repeated sampling, offers a way to interrogate these artifacts directly rather than speculate about them.

The stakes are high because the geroscience field has been waiting for credible endpoints. Interventions such as rapamycin, senolytic drugs and metformin are being tested in humans with the aim of extending healthspan, yet aging itself is not an approved trial endpoint. The TARGET trial of dasatinib and quercetin, reported by Justice and colleagues in Geroscience in 2018, was an early attempt to pair a senolytic intervention with biomarker readouts, and Moqri and colleagues laid out a systematic framework for biomarker validation in a 2023 review in Cell. What has been missing is the bridge between those frameworks and actual regulatory-grade evidence. A clock that reliably detects decelerated aging in a trial would transform the economics of the field, because trials measuring aging directly could be far shorter and smaller than trials waiting for fractures, infections or deaths to accumulate.

The commentary’s sober conclusion, that the trial taught the field more about the clocks than about the drug, should be read as progress rather than disappointment. First-in-class instruments are always refined through use. The trial provided longitudinal, randomized data against which clock behavior could be benchmarked, something no retrospective cohort can offer. It likely revealed issues of test-retest reliability, sensitivity floors and the influence of acute physiological fluctuations on plasma proteins, all of which are invisible in cross-sectional training data. Robertson and Chandra suggest that this kind of head-to-head deployment, with multiple clocks applied to the same trial samples, is exactly what the field needs to converge on which proteomic signatures are robust enough to carry forward toward validation as surrogate endpoints.

The path forward, as the commentary implies, will require standardization and skepticism in equal measure. Proteomic platforms need consistent calibration across laboratories; clocks need pre-registration and locked algorithms before trials begin, so that researchers cannot select the readout that happens to favor their result; and the field needs studies that test clocks against hard clinical outcomes prospectively, not just against chronological age. The 2026 trial of Zhavoronkov and colleagues will likely be remembered less for any verdict on its experimental therapy and more as the moment proteomic aging clocks left the training ground. If the lessons learned here are absorbed, the next generation of clocks may enter trials not as optimistic proxies but as instruments with documented performance characteristics, bringing the dream of measuring, and eventually modulating, human aging a measurable step closer.

Subject of Research: Validation of proteomic aging clocks as biomarkers in a clinical trial

Article Title: Proteomic aging clocks go to trial

Article References: Robertson, N. A., & Chandra, T. (2026). Proteomic aging clocks go to trial. Nature Biotechnology. https://doi.org/10.1038/s41587-026-03361-4

Image Credits: AI Generated

DOI: 10.1038/s41587-026-03361-4

Keywords: proteomic aging clocks, biological age, biomarkers, clinical trial, geroscience, plasma proteomics, surrogate endpoints, epigenetic clocks, longevity research, Nature Biotechnology, age prediction, biomarker validation

The research team, led by investigators at Yale University including Daniel Borrus and Albert Higgins-Chen, set out to answer a deceptively simple question: when should you actually trust an epigenetic clock result? The problem stems from the sheer number of clock models now in circulation. First-generation clocks such as those developed by Hannum and Horvath were trained to predict chronological age from DNA methylation patterns at CpG sites. Second-generation clocks like PhenoAge and GrimAge were trained instead on mortality risk, while pace-of-aging measures such as DunedinPoAm38 and DunedinPACE were built to capture how quickly a person is aging. Each clock reads a different slice of the methylome, and they do not always agree.

That disagreement matters because a single significant result from one clock can be produced by three very different things: a real, systemic shift in biological aging; a clock-specific effect tied to that particular model’s training data; or plain technical noise. The researchers’ central hypothesis was that a genuine change in biological aging should be visible across multiple clock models, while a lone significant clock is more likely to reflect noise or model quirks. To test this, they assembled twelve clock models, including high-reliability principal component versions of the classics, and re-analyzed longitudinal DNA methylation datasets spanning interventions, stressful events, and control cohorts.

The first target was the published intervention literature itself. The team identified three datasets that had previously reported a significant drop in biological age on a single clock, including an eight-week diet, exercise, and supplement program and a two-year folic acid and vitamin B12 trial, both originally scored with the Horvath multi-tissue clock. When they computed eleven additional clocks on the same data, not one changed significantly in the same direction as the original finding. In one dataset, ten of the eleven additional clocks actually trended in the opposite direction. The sobering implication is that some headline-grabbing claims of epigenetic age reversal may not survive contact with a broader panel of clocks.

To establish what a trustworthy signal looks like, the researchers turned the problem on its head and examined positive controls: events that should accelerate biological aging. They analyzed longitudinal datasets capturing intensive surgery, radiotherapy and chemotherapy for breast cancer, and SARS-CoV-2 infection. If clocks can detect a decrease in biological age, they should certainly detect an increase after a major physiological insult. That is exactly what they found, but with a telling pattern. In all three datasets, at least four of the twelve clocks registered significant age acceleration, and three mortality-linked measures, PCPhenoAge, PCGrimAge, and DunedinPACE, increased significantly in every single one.

From these positive controls, the team distilled three criteria for confidence in an epigenetic clock result. First, the reliable PC-based clocks should respond, since their principal component architecture suppresses the noise from individual CpGs. Second, the mortality-trained and pace-of-aging clocks should respond, because interventions target health outcomes rather than the passage of calendar time. Third, multiple clocks should move in the same direction, reflecting the shared aging signal they all carry. Notably, the original first-generation chronological clocks, Hannum, Horvath multi-tissue, and Horvath skin-and-blood, failed to detect a significant increase in any of the three stressor datasets, and the skin-and-blood clock even suggested a significant decrease after chemo-radiation, the opposite of what was biologically happening.

Why do reliable clocks outperform? The answer lies in a statistical property called the intraclass correlation coefficient, or ICC, which measures test-retest consistency. Earlier work had shown that re-testing the same person with standard clocks can produce swings of up to 2.4 years, while PC clocks cut that deviation to between 0.3 and 0.8 years. The team ran simulation studies of hypothetical clinical trials with 200 virtual participants and a known true effect, showing that high-ICC clocks detect genuine changes at much smaller effect sizes, while noisy clocks drown real signals in measurement error. Longitudinal control datasets confirmed the corollary: changes in reliable clocks correlate strongly with one another, with a mean pairwise correlation of 0.61 among reliable clocks versus just 0.20 among unreliable ones, meaning noisy clocks drift in spurious directions while reliable clocks move together.

Armed with these criteria, the researchers hunted for intervention datasets that passed all three tests, and found three. A two-year Mediterranean diet trial with two arms, one standard and one a green Mediterranean diet enriched with plants and polyphenols, and a three-week alcohol abstinence program in chronic drinkers all showed significant decreases in at least five clocks, with seven responding in the alcohol and green-Mediterranean datasets. Across these true-positive interventions, fourteen of eighteen reliable clock tests reached significance compared with five of eighteen for unreliable clocks, and the per-subject average across all twelve clocks fell significantly in every case. That twelve-clock average, the authors suggest, is not a new biomarker but a simple, interpretable filter for whether the clocks as a group agree.

The study also probes a subtler confound: cell composition. Blood methylation measurements reflect not just the aging of cells but shifts in the mixture of cell types present, and the team used EpiDISH deconvolution to adjust for twelve immune cell fractions. The results were strikingly heterogeneous. Surgery and chemo-radiotherapy effects shrank by roughly three-quarters after adjustment, indicating those signals were largely driven by compositional changes, while the SARS-CoV-2 effect was essentially unchanged and the Mediterranean diet arms retained most of their effect. The authors caution that this does not invalidate clocks trained on bulk blood, whose predictive power partly depends on cell fractions, but it does mean researchers should report cell-composition adjustments to interpret what a clock change actually represents.

The authors are candid about limitations: no intervention has yet been definitively proven to extend human lifespan, so there is no gold-standard positive control for rejuvenation, and a robust clock change does not automatically translate into added years of life. Still, the practical prescription is clear. Intervention studies should report multiple clocks, favor reliable PC versions and mortality- or pace-trained models, compute a per-subject clock average, and treat any solitary significant result with suspicion until it is replicated or supported by mechanistic evidence. As epigenetic clocks flood from the laboratory into clinics and consumer testing, this framework offers the field something it badly needs: a shared rulebook for separating genuine biological signal from statistical noise, before the anti-aging literature fills up with findings that cannot be repeated.

Subject of Research: Validation criteria for epigenetic clocks as biomarkers of aging interventions

Article Title: When to Trust Epigenetic Clocks: Avoiding False Positives in Aging Interventions

Article References: Borrus, D. S., Sehgal, R., Armstrong, J. F., Kasamoto, J., Gonzalez, J., & Higgins‐Chen, A. (2026). When to Trust Epigenetic Clocks: Avoiding False Positives in Aging Interventions. Aging Cell, 25(10), Article e70756. https://doi.org/10.1111/acel.70756

Image Credits: AI Generated

DOI: 10.1111/acel.70756

Keywords: epigenetic clocks, biological age, DNA methylation, aging biomarkers, longevity, DunedinPACE, GrimAge, PhenoAge, clinical trials, publication bias, anti-aging interventions, reliability

The eye is, in many respects, an ideal organ for aging research. It is paired, which allows within-person comparisons; it can be imaged repeatedly and non-invasively over decades; it can be treated locally with injections, lasers, and surgery; and its function can be quantified with extraordinary precision, from visual acuity to contrast sensitivity to dark adaptation. This tractability has fueled an explosion of models that estimate biological age from retinal images and other ocular data. Deep learning applied to retinal photographs has been reported to predict biological age and stratify morbidity and mortality risk. Multimodal retinal aging clocks built on optical coherence tomography and fundus imaging have been proposed for systemic health assessment. Liquid-biopsy proteomics of ocular fluids, combined with machine learning, has been used to identify cellular drivers of eye aging in vivo. Each of these approaches produces a number expressed in years, but the review emphasizes that the familiar unit conceals fundamental differences in what was measured and what the model was built to predict.

To untangle these differences, Bair proposes an eye-specific framework organized around five declarations that should accompany any ocular age estimate. The first is the proximal source: which biological material or signal the estimate derives from, whether a two-dimensional fundus image, a three-dimensional OCT scan, tear fluid or aqueous humor proteomics, or systemic omics that only indirectly reflect the eye. The second is the development target: what the model was actually trained to predict, which may be chronological age, a clinical outcome, or some composite index. The third is the reported quantity: whether the output is an absolute age, a residual difference from chronological age, an age-adjusted deviation, or a risk score. The fourth is the evidence design: whether the supporting study was cross-sectional, longitudinal, or interventional. The fifth is the intended use: whether the estimate is meant for mechanistic research, risk stratification, monitoring of an intervention, or clinical decision-making. Only when all five are stated explicitly, the review argues, can readers compare estimates across studies or translate them into practice.

Beyond these five declarations, the framework identifies several annotations that remain essential for any meaningful interpretation. Laterality matters, because the two eyes of one person are correlated and statistical approaches that treat eyes as independent observations inflate certainty; ophthalmology has long wrestled with this people-and-eyes problem. The reference population matters, because a model trained on one demographic or ethnic group may not calibrate to another. Disease and treatment context matters, since conditions such as diabetes, glaucoma, and age-related macular degeneration, and treatments such as intravitreal injections, alter the very features the models read. The technical version of the model matters, because updated algorithms can shift outputs without any biological change in the patient. And uncertainty in the stated output matters, because a single point estimate in years carries no information about its own reliability.

Perhaps the most consequential distinction the review draws is between a cross-sectional residual and a longitudinal aging rate. When a model estimates that a 60-year-old’s retina looks like that of a typical 66-year-old, the resulting six-year gap is a cross-sectional residual: a snapshot comparison against a reference population at one moment in time. It does not follow that the person is aging six years faster than everyone else. Repeated application of a cross-sectional model to the same individual over time does not establish an individual aging rate, because the model was never designed or validated to track change within a person. Longitudinal claims require repeated measurements evaluated against a longitudinal reference, and the observed change must be distinguishable from measurement variability. The review points to reference change values, a concept from clinical chemistry, as the appropriate tool for deciding whether a difference between two measurements on the same person exceeds the noise inherent in the method.

This confusion between cross-sectional and longitudinal inference is not unique to the eye. The brain-age literature, which pioneered the estimation of biological age from magnetic resonance imaging, has documented similar pitfalls, including regression dilution and the statistical traps that arise when age gaps are used as predictors of outcomes. The review imports lessons from that field, along with normative-modeling approaches that conceptualize disease as a deviation from expected trajectories rather than a categorical state. It also draws on the formal biomarker validation literature, including the framework that distinguishes discovery, analytical validation, clinical validation, and clinical utility, and on criteria for surrogate endpoints that specify when a biomarker can stand in for a clinical outcome in trials. By these standards, most ocular age estimates today sit firmly at the discovery stage.

The review further insists that visual-aging claims need validation against function, not just against images. The visual system offers psychophysical and performance-based measures, such as contrast sensitivity, rod-mediated dark adaptation, and hazard detection in night-driving simulators, as well as patient-reported outcomes. Delayed dark adaptation, for example, has been established as a functional biomarker for incident early age-related macular degeneration, and natural-history studies are now charting how visual function changes over years in normal aging and in early disease. An ocular age estimate that correlates with none of these functional measures risks being a statistical artifact rather than a biological signal. Differences across measurement domains, whether between an imaging-based clock and a proteomic one, must be assessed both against measurement error and against the distribution expected in the reference population, so that a difference is not declared meaningful simply because it is statistically detectable.

Even when a persistent difference survives these scrutiny layers, the review counsels restraint in interpretation. A residual age gap may support a candidate biological pattern only after technical explanations, such as image quality, camera differences, and algorithm version, and clinical explanations, such as occult disease and medication effects, have been evaluated and excluded. Only then does the gap become a hypothesis about retinal aging that can be tested in appropriate designs. Prognostic claims, in turn, require separate prospective evidence demonstrating that the estimate predicts outcomes in new samples, and clinical claims require decision-level evidence showing that acting on the estimate improves patient-relevant results. Tools such as decision curve analysis, which quantifies the net benefit of acting on a prediction at various threshold probabilities, and reporting standards such as TRIPOD+AI and its risk-of-bias companion PROBAST+AI, provide the methodological scaffolding for this progression from prediction to practice.

The practical upshot for clinicians, researchers, and the growing consumer market for biological-age tests is a checklist of questions to ask before taking an ocular age number at face value. What tissue or signal was measured? What was the model trained to predict, and against what reference population? Is the reported quantity an absolute age, a residual, or a calibrated deviation, and what is its uncertainty? Was the supporting evidence cross-sectional or longitudinal, and if longitudinal, was the change larger than measurement noise? Has the estimate been validated against visual function or patient-reported outcomes? And has any study shown that using the estimate changes decisions in ways that help patients? Until these questions are answered, the review concludes, a cross-sectional age residual should not be mistaken for an aging rate, a prognostic marker, or a clinically actionable test. The eye may indeed hold a clock, but reading it responsibly requires knowing what kind of clock it is, what it was built to measure, and what evidence stands behind each tick.

Subject of Research: Interpretation and validation of biological age estimates derived from ocular imaging and biomarkers

Article Title: Interpreting ocular age estimates: source, output, evidence, and clinical meaning

Article References: Bair, H. (2026). Interpreting ocular age estimates: source, output, evidence, and clinical meaning. Biogerontology, 27(5), Article 175. https://doi.org/10.1007/s10522-026-10525-x

Image Credits: AI Generated

DOI: 10.1007/s10522-026-10525-x

Keywords: biological age, ocular aging, retinal imaging, fundus photography, optical coherence tomography, biomarkers, deep learning, prediction models, longitudinal studies, visual function, biogerontology, clinical validation

The centerpiece of the study is a new metric called the Empirical Dietary Index for Slower Epigenetic Aging, or EDISEA. Unlike traditional dietary scores, which are typically constructed from hypotheses about which foods benefit health, EDISEA is data-driven. The researchers began with a derivation sample of 1,711 participants and searched for the dietary features that best predicted lower acceleration of GrimAge2, a second-generation epigenetic clock widely regarded as one of the strongest molecular predictors of mortality and morbidity. Epigenetic clocks work by measuring patterns of DNA methylation, the chemical tags that attach to DNA and influence gene activity without changing the underlying genetic code. As people age, these methylation patterns shift in predictable ways, allowing scientists to estimate biological age from a blood sample with remarkable precision. When a person’s epigenetic age exceeds their chronological age, that gap, known as age acceleration, signals faster biological wear and tear.

The technical elegance of this approach lies in its directionality. Most prior studies asked whether existing diet scores correlate with epigenetic aging; this team inverted the question, asking which foods and nutrients the aging clocks themselves point toward. The resulting index is empirical, meaning it emerges from observed associations in real population data rather than from expert opinion. In the derivation sample, participants with the highest EDISEA scores showed epigenetic aging that was 4.17 years slower than expected, a difference that was highly statistically significant with a P-value below 0.001. Crucially, the finding was not an artifact of the sample used to build the index. In an independent validation sample of 2,311 participants, higher EDISEA was associated with 3.77 years slower epigenetic aging, confirming that the pattern generalizes beyond the population in which it was developed.

Slower epigenetic aging is an intriguing molecular signal, but the question that matters for public health is whether it translates into longer, healthier lives. To address this, the researchers turned to outcome-wide association analyses across multiple large cohorts, including the Health and Retirement Study, the National Health and Nutrition Examination Survey, the English Longitudinal Study of Ageing, and UK Biobank. The mortality results were consistent and substantial: across two US cohorts, higher EDISEA was associated with a pooled hazard ratio of 0.80 for all-cause mortality, with a 95 percent confidence interval of 0.76 to 0.85. In practical terms, that corresponds to roughly a 20 percent lower risk of death from any cause among people whose diets most closely matched the epigenetic aging-informed pattern, compared with those whose diets matched it least.

The associations extended well beyond survival. Higher EDISEA scores were linked to lower risks of a range of aging-related diseases, fewer functional limitations, and reduced multimorbidity, the co-occurrence of multiple chronic conditions that places the greatest burden on aging populations and health systems. Multimorbidity is one of the most challenging features of human aging because it resists single-disease interventions; a dietary pattern that appears to shift the entire trajectory of aging could, in principle, delay many conditions at once. The breadth of these findings across both US and UK aging cohorts strengthens the case that the pattern captures something fundamental about how nutrition interacts with the biology of growing older.

Perhaps the most visually compelling evidence comes from the molecular and neuroimaging analyses conducted in a large UK population-based cohort. Participants with higher EDISEA scores showed more favorable profiles across vascular, inflammatory, metabolic, and brain structural measures. Neuroimaging signatures are particularly noteworthy because brain structure changes slowly and is difficult to modify through intervention; observing that an epigenetic aging-informed diet is associated with healthier brain structural profiles suggests that the dietary pattern may be tracking genuine biological preservation rather than merely reflecting healthier lifestyles in general. The convergence of molecular blood markers, imaging data, and clinical outcomes provides a multi-layered picture in which diet, epigenetic aging, and downstream health all move together in a coherent direction.

The study’s design deserves careful attention when weighing what these results mean. Because the analyses are observational, they demonstrate association rather than causation. People who eat in ways that slow their epigenetic clocks may also differ in other respects, such as socioeconomic circumstances, physical activity, smoking habits, or access to health care, and the researchers used statistical adjustment across large cohorts to account for many such confounders. Still, residual confounding cannot be fully excluded, and randomized trials will ultimately be needed to establish whether adopting the EDISEA pattern directly causes slower epigenetic aging and better health outcomes. What distinguishes this work from much of nutritional epidemiology, however, is its explicit validation architecture: the index was derived in one sample, tested in an independent sample, and then evaluated against hard outcomes including mortality in multiple additional cohorts spanning two countries.

The implications for research and clinical practice could be considerable. Existing dietary guidelines are built largely around cardiovascular and metabolic endpoints, and indices designed for those outcomes may not be optimal for targeting the aging process itself. EDISEA offers an epigenetic aging-informed framework that researchers can use to study dietary strategies relevant to healthy aging, and it provides a template for a new generation of dietary indices built around molecular biomarkers rather than disease diagnoses alone. If future interventional studies confirm the observational findings, clinicians could eventually use epigenetic clocks both to identify individuals whose biological aging is outpacing their years and to tailor dietary recommendations aimed at slowing that process. The approach also dovetails with a broader movement in geroscience, which seeks to treat aging as a modifiable risk factor underlying many chronic diseases simultaneously.

There are important caveats to keep in view. Epigenetic clocks continue to evolve, and GrimAge2, while among the best-validated predictors, is one of several competing measures of biological age. The study populations were drawn from the United States and the United Kingdom, and dietary patterns, food environments, and population genetics differ across the world, so generalizability to other settings remains to be demonstrated. The index is also empirical, which means its specific food components reflect the populations used to derive it and may be refined as more data accumulate. The authors note that the work is intended as a framework for studying dietary strategies rather than a prescription, and the open-access publication, released on 25 September 2026 with full supplementary data and a transparent peer-review file, allows other research groups to scrutinize and extend the analysis.

Even with those qualifications, the study represents a conceptual milestone. It demonstrates that the molecular machinery of aging can serve as a design target for nutrition science, and that a diet built around that target is associated with slower epigenetic aging, lower mortality, less disease, and healthier brains and blood vessels across thousands of people in multiple nations. As epigenetic testing becomes faster and cheaper, and as the science of biological age advances from the laboratory toward the clinic, the idea that what we eat can be tuned to the ticking of our methylation clocks may move from an intriguing correlation to a practical strategy for extending not just lifespan, but healthspan, the years of life spent free of disease and disability. For now, the message is one of cautious excitement: a data-driven eating pattern informed by epigenetic aging is associated with a spectrum of better aging outcomes, and the research community now has both the tool and the motivation to test whether that association can be turned into cause and effect.

Subject of Research: An empirically derived dietary index targeting epigenetic aging and its associations with aging-related health outcomes

Article Title: An epigenetic aging-informed dietary pattern is associated with a spectrum of aging-related health outcomes

Article References: Lai, S., Yu, J., Zhang, L., Wu, M., Peng, G., Zong, G., Ma, H., Yuan, C., Luo, B., & Chen, H. (2026). An epigenetic aging-informed dietary pattern is associated with a spectrum of aging-related health outcomes. Nature Communications. https://doi.org/10.1038/s41467-026-77790-9

Image Credits: AI Generated

DOI: 10.1038/s41467-026-77790-9

Keywords: epigenetic aging, DNA methylation, GrimAge2, dietary patterns, EDISEA, biological age, healthy aging, mortality, multimorbidity, neuroimaging, nutritional epidemiology, Nature Communications

The motivation is demographic as much as scientific. The United States now counts nearly 60 million residents aged 65 and older, a figure projected to grow by another 42 percent by 2050, when more than 23 percent of Americans will be over 65. Age remains the single strongest risk factor for chronic disease, and for many older adults the burden is compounded by cognitive decline, including Alzheimer’s disease and related dementias. The researchers argue that promoting healthy aging and blunting the impact of age-related disease has become a major national social and economic challenge, one that demands biobehavioral data detailed enough to guide real public health interventions rather than isolated correlations.

To build that dataset, ASHA convenes experts spanning biomedical imaging, integrative physiology, biomechanics and exercise science, neuroscience, behavioral science, social work, public health, and biostatistics. Each participant completes two in-person visits lasting five to eight hours each. The protocol begins with informed consent and questionnaires delivered through Vibrent Health, a secure, HIPAA-compliant online platform that lets people complete medical histories remotely. Those histories cover medications, surgeries, physical activity, smoking, alcohol consumption, menstruation history, fall history, education, marital status, and even zip code, painting a social and environmental backdrop against which the biological measurements can later be interpreted.

The imaging centerpiece is a whole-body MRI protocol performed on a 3 Tesla Siemens MAGNETOM Vida scanner at the university’s Clinical Imaging Research Center, operated by a licensed radiologic technologist with breaks built in for comfort. The brain portion stacks structural and functional sequences: high-resolution T1-weighted MPRAGE anatomy, T2-weighted FLAIR to flag white matter lesions, and diffusion tensor imaging to probe the integrity of neural wiring. Cerebral blood flow is measured at rest with arterial spin labeling, while resting-state and task-based BOLD functional MRI, recorded alongside end-tidal CO2, quantify intrinsic brain networks, cerebrovascular reactivity, and activation during language, motor, and executive function tasks such as picture naming and finger tapping.

The rest of the body gets equal scrutiny. A head-to-toe stack of DIXON images separates water and fat signals to map skeletal muscle size and fat distribution across every anatomical region, while a multi-echo variant quantifies liver fat and iron, key markers for non-alcoholic fatty liver disease and iron overload. Thigh muscle imaging adds quantitative T1, T2, and fat-fraction mapping to detect intramuscular fat infiltration, a hallmark of muscle degeneration. Cardiac MRI rounds out the session with cine imaging of ventricular function, tissue tagging to measure strain, T1 and T2 mapping for fibrosis and edema, diffusion tensor imaging of heart muscle architecture, aortic distensibility and pulse wave velocity as markers of vascular stiffness, 4D flow imaging of aortic hemodynamics, and oxygen-sensitive sequences that probe coronary endothelial function without ever threading a catheter.

On a separate day, after a six-hour fast, participants undergo a vascular battery that spans macro- and microcirculation. Blood pressure is measured sitting, standing, and supine to assess postural regulation; ankle-brachial and toe-brachial indices screen for peripheral artery disease; and the SphygmoCor system derives central pulse wave analysis and pulse wave velocity. Beat-to-beat pressure is captured from a finger cuff via photoplethysmography, and brachial artery flow-mediated dilation, the standard noninvasive test of endothelial health, is measured with Doppler ultrasound and rapid cuff inflation. Near-infrared diffuse correlation spectroscopy over the forearm adds a window into microvascular reactivity and tissue oxygenation kinetics, complementing the larger-vessel measures.

Following an approximately 80-milliliter blood draw, part of which goes to clinical testing and the rest to a biorepository of PAXgene tubes and frozen plasma and serum for future genetic and transcriptomic work, participants refuel with a snack and face a cognitive battery. It includes the Montreal Cognitive Assessment for global screening, the NIH Toolbox Cognition Battery and an odor identification test for domain-specific evaluation, and a prospective memory test probing the ability to act on future intentions cued by the environment. Physical function testing then begins with the Short Physical Performance Battery and grip strength dynamometry, followed by gait analysis on an instrumented ZenoMat walkway under both normal conditions and cognitive load, counting backward by ones or sevens or naming animals while walking. A six-minute walk test and a DEXA scan for lean mass, fat mass, and bone density close the visit.

Between clinic visits, the study follows participants into daily life. For seven days they wear a Fitbit Inspire 3 tracking steps, activity intensity, and sleep stages; a cuffless Hilo bracelet, formerly known as Aktiia, that captures blood pressure fluctuations around the clock; and a FreeStyle Libre continuous glucose monitor sampling interstitial glucose every fifteen minutes. Together these devices record glycemic variability, postprandial responses, nocturnal glucose trends, and autonomic cardiovascular rhythms in real-world settings, adding ecological validity that clinic snapshots cannot provide. The design also builds in safeguards: incidental MRI findings are reviewed by the imaging center’s medical director and communicated to participants with recommendations for follow-up, and the study itself offers no diagnosis or treatment.

What may set ASHA apart most is its deliberate diversity and equity focus. The 600 participants, recruited beginning in November 2024 with completion expected by late 2028, will be evenly distributed across Non-Hispanic Black, Non-Hispanic White, Hispanic or Latino, and Asian residents of Tarrant County, one of the largest and most demographically diverse counties in the country, with balanced sex representation in each group. Concentrating recruitment in a single county trades some generalizability for rigor: the same instruments and personnel assess every participant, minimizing measurement noise. Exclusions screen out active cancer treatment, advanced chronic disease, dementia, non-ambulatory status, severe gait disorders, and other conditions that would confound the healthy-aging signal. The investigators acknowledge the single-site limitation and plan controlled-access data sharing so that qualified researchers elsewhere can test whether the patterns hold. Ultimately, by measuring organ systems in concert rather than isolation, the team hopes to build biological age metrics, from frailty indices to imaging-derived biomarkers, that reveal why some people age with resilience while others decline, and to identify the modifiable factors that keep more of us independent for longer.

Subject of Research: A multidisciplinary community cohort study investigating biological, psychological, and social mechanisms of age-related functional decline through multi-organ imaging and monitoring.

Article Title: The Arlington study of healthy aging: a multidisciplinary community cohort study of functional decline with age

Article References: Davis, D. L., Shah, R., Pixler, L., Mai, D. M., Santos, J., Chandler, C., White, K., Nguyen, J., Moradi, A. S., Choudhari, J., Arena-Marshall, C., Johnson, C., Kamel, L., Liao, Y., Greer, T. L., Cooper, C., Ball, H., Fields, N. L., Wang, X., … Nelson, M. D. (2026). The Arlington study of healthy aging: a multidisciplinary community cohort study of functional decline with age. GeroScience, 48(5), 6409-6420. https://doi.org/10.1007/s11357-026-02498-z

Image Credits: AI Generated

DOI: 10.1007/s11357-026-02498-z

Keywords: healthy aging, cohort study, whole-body MRI, functional decline, vascular function, cognitive aging, wearable monitoring, sarcopenia, biological age, older adults, geroscience, health equity