Pith. sign in

REVIEW 2 major objections 5 minor 4 references

A self-supervised model trained on 11,540 DXA scans extracts hidden multi-system disease risk and biological age from routine bone scans.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A JEPA-based vision model trained on raw DXA body scans predicts disease risk, biological age, and mortality better than standard summary DXA features on several key endpoints.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A genuinely new SSL-for-DXA result with real external validation; the incident-disease and mortality claims need a center-adjusted sensitivity check before they fully convince. the 2 major comments →

arxiv 2608.02208 v1 pith:UQ7U4TXX submitted 2026-08-03 cs.CV q-bio.QM

Self-supervised DXA representations encode multi-system disease risk, biological aging and heritability

Gil Sasson , Zachary Levine , Smadar Shilo , Sarah Kohn , Guy Lutsker , Anastasia Godneva , Adam Gabet , David Krongauz
show 4 more authors
Adina Weinberger Yann LeCun Randall Balestriero Eran Segal
This is my paper
classification cs.CV q-bio.QM
keywords self-supervised learningDXAJEPAbody compositionbiological agedisease riskheritabilitygenome-wide association
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that ordinary whole-body DXA scans, already taken for bone-density screening, contain systemic health information that standard scalar readouts throw away — and that a self-supervised model can recover that information without any disease labels. The model, LeDXA, is trained from scratch on 11,540 unlabeled scans using a joint-embedding predictive architecture (JEPA), then frozen and probed in an independent cohort of 47,400 scans. The authors report that its embeddings predict prevalent and incident disease, chronological age, and a biological-age gap that tracks mortality, and that they carry heritable genetic signal matching known body-composition loci. A sympathetic reader would care because, if true, routine DXA imaging could become a multi-system risk-stratification and aging-monitoring tool at almost no extra clinical cost.

Core claim

LeDXA claims that the spatial geometry of a DXA scan — the distribution of bone, lean tissue, and fat across the body — encodes physiological state beyond the regional sums currently reported. Trained from random initialization on internal scans with the LeJEPA objective (predicting latent global-view representations rather than pixels), the frozen encoder recovers scanner-derived body composition with r ≈ 0.96, predicts physiological biomarkers better than a general-purpose vision model, improves incident-disease discrimination over tabular DXA measures in 9 of 20 external endpoints, and yields a biological-age gap with a 45% higher mortality hazard in the oldest-appearing quartile. The aut

What carries the argument

The central object is the joint-embedding predictive architecture (JEPA) applied to DXA: a ViT-Small/16 encoder that processes two per-scan image channels (bone and soft tissue), each expanded into 2 global and 8 local views. Every view is trained to predict the centroid of the global-view projections in a 64-dimensional latent space, with a sketched isotropic Gaussian regularization (SIGReg) preventing collapse. This objective — predicting latent structure rather than reconstructing pixels — is what the paper argues lets a compact model learn whole-body anatomical shape from only 11,540 scans and transfer to an unseen cohort.

Load-bearing premise

The load-bearing premise is that the frozen representations trained on one scanner family (GE Lunar Prodigy) transfer to another (GE Lunar iDXA) without encoding scanner- or site-specific technical variance that correlates with health outcomes; the paper does not test for scanner or site effects in the embeddings.

What would settle it

If embeddings of the same person acquired on different scanner models are strongly separated by scanner identity, or if biological-age-gap and disease associations vanish when the analysis is restricted to a single scanner model or adjusted for assessment center, the central claim of a systemic health signal would be undermined.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Routine DXA scans could serve as an inexpensive, label-free source of multi-system risk markers, with LeDXA improving incident-disease discrimination over scanner readouts in 9 of 20 tested endpoints.
  • On the largest gains, hip arthrosis, LeDXA flagged 66% of incident cases in its top risk quartile versus 41% for tabular measures by end of follow-up; knee arthrosis and type 2 diabetes showed similar advantages.
  • The biological-age gap — a residual from age prediction — stratified all-cause mortality independently of age and sex, with a 45% higher hazard in the oldest-appearing quartile and monotonic ordering across quartiles.
  • The gap moved in paired before/after analyses: it decreased after starting hormone-replacement therapy in women and antidepressants in men, suggesting the representation is sensitive to modifiable physiological change.
  • Because the embeddings are heritable (mean SNP-h² 0.143, above a general-purpose model's 0.098) and recover known loci without genetic training, image-derived phenotypes could be used as quantitative traits in future genome-wide studies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: if these findings replicate in clinical and ancestry-diverse populations, DXA-based biological age could become an endpoint for monitoring interventions aimed at preserving lean mass and bone during weight-loss therapy — a use the authors gesture at but do not test directly.
  • Beyond the paper: the unsupervised female cluster analysis suggests the embedding separates a healthy body-composition axis at matched age and BMI; a natural next test is whether these clusters are stable across scanner geometries and whether they predict future disease independent of standard risk factors.
  • Beyond the paper: the authors implicitly assume that scanner- and site-specific technical variation does not drive the cross-cohort signal; an explicit test would be to evaluate embeddings within a single scanner model or with center identity as a covariate.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper introduces LeDXA, a compact ViT-Small/JEPA model trained from scratch on 11,540 HPP whole-body DXA scans, and evaluates its frozen embeddings on 47,400 external UK Biobank scans. It reports that LeDXA recovers scanner-derived body-composition and BMD readouts, improves prevalent- and incident-disease discrimination over a DXA-tabular baseline and over DINOv3 for a subset of endpoints (largest gains for hip/knee arthrosis and type 2 diabetes), produces a biological-age gap that tracks disease burden and mortality, and yields embedding PCs whose GWAS recover known body-composition and bone loci and that are more heritable than DINOv3's. The central claim is that raw DXA images contain systemic health signal beyond the scalar readouts currently used clinically.

Significance. The study is well-powered, externally validated with a frozen model, and benchmarked against a strong tabular baseline and a general-purpose vision model. The GWAS and SNP-heritability analyses are a particularly credible way to connect learned representations to biology, and the public code/data availability is a strength. If the scanner/site confounding concern is resolved, the result would be valuable: it would show that a relatively small, domain-specific SSL model can extract prognostic and heritable body-composition representations from routine DXA scans with modest compute. The unsupervised cluster analysis, with matched age/BMI and independent omics validation, is a further strength that does not depend on the longitudinal claims.

major comments (2)
  1. [Methods: Data acquisition and preprocessing; UKBB analyses (Fig. 3, Table S5)] The central 'beyond standard DXA readouts' claim depends on the cross-cohort gains being biological rather than technical. HPP scans are GE Lunar Prodigy; UKBB scans are GE Lunar iDXA acquired at multiple assessment centers. The DXA-tabular baseline explicitly excludes scanner details, measurement dates, and assessment-center identifiers (Methods: DXA tabular features), but the image embeddings are not purged of acquisition properties, and none of the prevalent-disease, incident-disease, biological-age, or mortality models adjust for assessment center or scanner software. No analysis in the paper quantifies how much LeDXA embedding variance is explained by center. Please provide a center-stratified or center-adjusted re-analysis of the incident-disease C-index gains in Table S5 and Fig. 3a,b and of the prevalent-disease AUROC comparisons, and report the association of LeDXA PCs with asse
  2. [Methods: All-cause mortality; Fig. 5e,f] The mortality analysis adjusts only for chronological age and sex. The biological-age gap is a body-composition residual, so BMI, smoking, comorbidity, and socioeconomic position are plausible confounders. The Q4-versus-Q1 HR of 1.45 and per-year HR of 1.06 should be re-estimated with adjustment for BMI, smoking, and prevalent disease (and for assessment center, as above). With 377 deaths over a median 3.4-year follow-up, the confidence intervals are wide, and without these adjustments the mortality gradient does not yet establish that the gap captures biological aging beyond standard risk factors.
minor comments (5)
  1. [Results: Prevalent disease; Tables S2/S3] LeDXA improves over the tabular baseline in 12/37 HPP and 9/28 UKBB endpoints, with no significant difference on most endpoints. The abstract's 'wide spectrum' should be calibrated to this minority-of-endpoints pattern.
  2. [Methods: Longitudinal incident-disease risk] The choice of the first 100 embedding PCs is not justified beyond a variance threshold. A sensitivity analysis varying the PC count would strengthen the C-index comparisons.
  3. [Methods: Biological age] The detrending polynomial degree (2) is a free parameter; report sensitivity to degree 3 and 4. Also, the negligible correlation of the gap with chronological age is by construction after OLS detrending and should not be presented as an empirical finding.
  4. [Table S7 / Fig. 5h] The HRT and antidepressant results are based on 11 and 17 paired participants, with self-reported medication timing, after testing 35 class-by-sex combinations. These should be framed as hypothesis-generating, not as evidence that the gap is modifiable.
  5. [Discussion: Limitations] The limitations paragraph acknowledges healthy-volunteer bias, ancestry, and short follow-up, but does not mention scanner/site differences. This should be added, along with the planned center-adjusted analysis or an explicit acknowledgment of the residual risk.

Circularity Check

1 steps flagged

No significant circularity: the central cross-cohort predictions rest on external UKBB validation and a frozen model; one minor self-definitional presentation in the age-gap residual.

specific steps
  1. self definitional [Methods, 'Biological age — Gap derivation'; also Results, 'LeDXA-derived biological age predicts clinical outcomes and tracks pharmacological responses']
    "The biological-age gap was defined as the residual from an ordinary least-squares model regressing out-of-fold predicted age on a degree-2 polynomial of chronological age (age and age²). This detrending removes both linear and quadratic age dependence, making the gap orthogonal to chronological age by construction. Empirically, the gap showed negligible correlation with chronological age overall (r = −0.002) and within each sex (females r = +0.006, males r = −0.019...)"

    The 'empirical' zero correlation with chronological age is not an independent empirical finding: OLS residuals are orthogonal to the regressors (age and age²) by construction. The paper presents this mathematical identity as an empirical check, then uses it to argue that the Q4 mortality gradient 'cannot be explained by older subjects accumulating in Q4.' That inference is valid, but it is a property of the residual definition, not independent evidence. This is a minor self-definitional step; it does not make the mortality or incident-disease results circular, because those associations are computed against external outcomes that were not used to define the gap.

full rationale

The paper's central claims are not circular. LeDXA was trained from scratch on HPP scans with a self-supervised objective that does not use any outcome labels, then frozen and applied to 47,400 external UK Biobank scans ('No UKBB data were used at any stage of pretraining; all UKBB analyses were performed on embeddings produced by the frozen HPP-trained model'). The main benchmarks compare LeDXA against scanner-derived tabular features and DINOv3 under covariate-adjusted, split-based evaluation; the incident-disease Cox models and mortality analyses are external-outcome tests, not fits of the embeddings to those outcomes. The GWAS and heritability analyses are independent genetic validations. The LeJEPA citation (ref 34, by co-authors) is a method citation and is not load-bearing for the empirical results; no uniqueness theorem is invoked. The only definitional artifact is the age-gap residual's orthogonality to chronological age, which is true by construction and is presented as an 'empirical' check; this is minor and does not affect the paper's overall external-validation logic. Scanner/site confounding in the UKBB embeddings is a plausible correctness risk, but it is not a circularity because the paper does not define its predictions in terms of the scanner metadata.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The paper introduces no new physical entities or forces. Its central empirical claims rest on domain assumptions: that LeJEPA training is stable on DXA; that 384x128 resizing preserves signal; that UKBB outcome coding is reliable; and that the age-gap residual reflects biology rather than scanner artifact. Free parameters are standard regularization/dimensionality choices.

free parameters (5)
  • Cox model L2 penalties = per endpoint, searched over {0.01, 0.1, 1, 10, 100}
    Selected on training folds for each incident-disease model; standard regularization tuning, not a scientific parameter.
  • Ridge/logistic C penalties = grid-selected (5 values)
    For linear probing and classification; tuned via nested cross-validation. Not central to the claim.
  • Number of embedding PCs for Cox models = 100 (90% LeDXA / 87% DINOv3 variance)
    Chosen to capture most variance; not tuned to outcomes.
  • Number of embedding PCs for GWAS = 20
    Top 20 PCs used as phenotypes; arbitrary but standard.
  • Biological-age detrending polynomial degree = 2 (age and age^2)
    Residuals after OLS on age and age^2 make the gap orthogonal to age; a modeling choice.
axioms (6)
  • domain assumption LeJEPA provides stable, non-collapsing SSL training without hand-crafted tricks.
    Used as the pretraining framework (ref 34); the paper does not prove this property for DXA, relying on the cited method.
  • ad hoc to paper The 384x128 resized DXA image preserves the clinically relevant spatial information.
    All scans resized to this fixed size; no ablation shows this resolution is sufficient.
  • domain assumption UK Biobank hospital-episode statistics and self-report provide accurate disease ascertainment.
    Used for prevalent and incident disease labels; coding errors could attenuate effects.
  • standard math LD score regression yields unbiased SNP heritability estimates for embedding PCs.
    Standard method; assumes LD reference panel and summary statistics are correct.
  • domain assumption The biological-age gap residual is orthogonal to chronological age and reflects biological aging, not technical artifact.
    Gap is residual after regression on age and age^2; paper shows it is uncorrelated with age and matches known aging patterns, but scanner/site effects are not excluded.
  • domain assumption DINOv3 is a representative state-of-the-art general-purpose vision model for comparison.
    Used as baseline; trained on natural images, not medical.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-supervised DXA representations encode multi-system disease risk, biological aging and heritability." pith.science (2026). https://pith.science/paper/UQ7U4TXX

@misc{pith2026260802208,
  author       = {Pith},
  title        = {Pith review of: Self-supervised DXA representations encode multi-system disease risk, biological aging and heritability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UQ7U4TXX}},
  note         = {Machine review of arXiv:2608.02208}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Whole-body dual-energy X-ray absorptiometry (DXA) scans are routinely acquired to measure bone density and regional body composition, leaving their spatial structure largely unused. Here, we show that self-supervised learning (SSL) can convert raw DXA images into representations of systemic health. We introduce LeDXA, a vision model based on a joint-embedding predictive architecture (JEPA) that learns by predicting latent representations rather than reconstructing pixels. Trained from scratch on 11,540 unlabeled Human Phenotype Project scans, LeDXA was evaluated internally and on 47,400 external UK Biobank (UKBB) scans. It improved cross-cohort prediction of prevalent diseases and biomarkers beyond scanner-derived DXA measurements and DINOv3, a state-of-the-art general-purpose model, despite approximately 150,000-fold fewer training images and nearly 40-fold fewer parameters. Over a median 4.3-year UKBB follow-up, LeDXA improved incident disease prediction over tabular DXA measures, with the largest gains for hip and knee arthrosis and type 2 diabetes. For hip arthrosis, 66% of incident cases occurred in the highest-risk quartile versus 41% for tabular measures. Its representations predicted chronological age externally (r = 0.88; mean absolute error = 2.90 years), and the biological-age gap tracked broader disease burden and a 45% higher mortality hazard in the oldest-appearing quartile. The gap also decreased in women after starting hormone-replacement therapy, suggesting it may be modifiable. Genome-wide associations recovered mostly known body-composition and bone-density loci, and LeDXA embeddings were more heritable than DINOv3's. These findings reveal prognostic information in DXA images that conventional readouts discard, learnable with relatively little data and modest compute.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

4 extracted references · 4 linked inside Pith

  1. [16]

    Azizi, S. et al. Robust and data-efficient generalization of self-supervised machine learning for diagnostic imaging. Nat. Biomed. Eng. 7, 756–779 (2023). 17. Chen, R. J. et al. Towards a general-purpose foundation model for computational pathology. Nat. Med. 30, 850–862 (2024). 18. Vorontsov, E. et al. Virchow: A Million-Slide Digital Pathology Foundatio...

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.07778 2023

  2. [31]

    Assran, M. et al. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture. Preprint at https://doi.org/10.48550/ARXIV.2301.08243 (2023). 32. Bardes, A. et al. Revisiting Feature Prediction for Learning Visual Representations from Video. Preprint at https://doi.org/10.48550/ARXIV.2404.08471 (2024). 33. Dong, Z. et al. Brain-JEPA...

    Pith/arXiv arXiv doi:10.48550/arxiv.2301.08243 2023

  3. [45]

    Cruz-Jentoft, A. J. et al. Sarcopenia: revised European consensus on definition and diagnosis. Age Ageing 48, 16–31 (2019). 46. Mwinyi, J. et al. Long-term changes in adiposity markers during and after antidepressant therapy in a community cohort. Transl. Psychiatry 14, 330 (2024). 47. Gambacciani, M. & Levancini, M. Featured Editorial Hormone replacement...

    Pith/arXiv arXiv doi:10.48550/arxiv.1802.03426 2019

  4. [60]

    & Mantzoros, C

    Karakasis, P., Patoulias, D., Fragakis, N. & Mantzoros, C. S. Effect of glucagon-like peptide-1 receptor agonists and co-agonists on body composition: Systematic review and network meta-analysis. Metabolism 164, 156113 (2025). 61. Wang, D. et al. GDF15 promotes weight loss by enhancing energy expenditure in muscle. Nature 619, 143–150 (2023). 62. Dosovits...

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.