Edition and source
This reading edition was reconstructed from the complete 62-page arXiv v1 PDF because that version has no native HTML. Original figure bitmaps, all supplementary tables, formula images and 73 bibliography entries are preserved. The notes explain the method and observational evidence; the original source prose is retained in the HTML cache.
Abstract
A DXA scan contains more information than its usual bone-density and body-composition totals. LeDXA learns a compact image vector without diagnosis labels, then trains separate predictors on those vectors. Its external-cohort results suggest useful risk information, especially for joint disease. The study is observational: prediction, an age-associated score and treatment-associated changes are not proofs of causation or a validated clinical decision tool.
Introduction
Think of a scan as a detailed map and the usual scanner readouts as a few totals from that map. LeDXA tries to retain useful spatial patterns. It uses LeJEPA: different crops of a scan should agree in representation space, while a distribution regularizer prevents every scan receiving the same vector. The reported data and parameter savings compare a domain-specific model with a much larger general image model; they do not establish a universal scaling advantage.

Results
The results progress from recovering familiar scan measurements to disease prediction, genetics, an age-associated score and exploratory clustering. The following sections explain what each evaluation can establish.
Study populations and DXA imaging
Training uses 11,540 scans from 8,820 Human Phenotype Project participants. The frozen encoder is evaluated on a separate UK Biobank cohort with 47,400 scans. Repeated scans and participants are different counting units. Cohort filtering and missing outcomes make the actual sample smaller for individual analyses; the external cohort contributes no encoder-pretraining images.
LeDXA predicts multi-system disease and physiological markers
The learned vectors recover conventional composition measurements and support age, biomarker and disease predictors. AUROC measures how well a classifier ranks cases above non-cases; 0.5 is chance and 1 is perfect ranking. LeDXA significantly beats tabular measurements for 12 of 37 internal and 9 of 28 external disease endpoints, with ties or occasional losses elsewhere. All feature-based disease models also receive age, sex and body mass index. These are supervised downstream tests of an initially self-supervised encoder.

Longitudinal risk prediction of incident disease
Here the scan precedes a recorded diagnosis. Cox models rank future event risk while accounting for different follow-up times. LeDXA improves over the tabular baseline on 9 of 20 tested endpoints; hip and knee arthrosis show the largest gains. Capturing 66% of hip cases in the highest-risk quarter describes case concentration, not 66% diagnostic accuracy or proof that screening improves outcomes. LeDXA and DINOv3 are similar on most incident endpoints.

LeDXA-derived representations match known genetic signals in an external cohort
A genome-wide association study asks whether common genetic variants correlate with components of the image vectors. Many matches overlap known body-composition and bone loci. LeDXA components show more estimated common-variant heritability than DINOv3 components. This supports biological alignment of some variation, while neither a statistical association nor mapping a variant to a nearby gene establishes a causal biological mechanism.

LeDXA-derived biological age predicts clinical outcomes and tracks pharmacological responses
A supervised age predictor is corrected for its systematic relationship with chronological age; its residual becomes the age gap. An older-looking scan is associated with greater disease burden and higher mortality hazard after the stated adjustments. The gap is a constructed score rather than a direct measurement of a universal biological clock. Medication comparisons are small, uncontrolled before/after analyses: 11 women for hormone replacement and 17 men for antidepressants. They cannot show that either treatment reverses aging.

Unsupervised LeDXA embeddings identify a lean- and high-bone-density female phenotype associated with favorable physiological profiles
The authors project female scan vectors to two dimensions and cluster them, then match the groups on age and body mass index. One group has more lean tissue and higher bone density, alongside several favorable measured traits. Matching reduces two specific confounders, but does not randomize the groups. Because clustering happens after a nonlinear two-dimensional projection, the visible groups also depend on the analysis pipeline.

Discussion
The engineering lesson is that a small model trained on the right imaging domain can preserve useful spatial information that scalar summaries lose. External-cohort evaluation is a strength. Important limits include selected volunteer cohorts, observational outcomes, self-reported diagnoses or medication histories, scanner differences and untested clinical utility. The embeddings are promising research features, with deployment and treatment decisions requiring additional evidence.
Methods
The methods distinguish encoder pretraining from supervised task fitting. Read this part to understand which labels, covariates, cohort filters and tuning choices enter each reported result.
Study populations
Internal participants come from an Israeli longitudinal cohort and external participants from UK Biobank. Eligibility, image completeness, covariates and outcome availability define each analysis set. Subject-level splits keep repeated scans from the same person out of both training and validation within a split; the findings still reflect the populations and scanners represented here.
Data acquisition and preprocessing
Bone and soft-tissue images are resized and replicated into three channels. Regional femur and lumbar scans supplement the full-body inputs for relevant tasks. This standardizes the input tensor while preserving the distinction between full-body and clinically targeted views.
Self-supervised pretraining with LeJEPA
A small vision transformer turns each crop into features, and a projection network supplies the training coordinates. Two global views and eight local crops of a scan are encouraged to agree. SIGReg pushes projected coordinates toward a standard Gaussian distribution so different examples cannot all collapse to the same point. This is crop-based representation learning, without future-state prediction or diagnosis labels in pretraining. The exact source equations remain below.
Equations



Training configuration
Training uses AdamW, 400 epochs, mixed precision and a single NVIDIA L40S graphics processor, taking approximately 10 hours under the stated setup. Once trained, the encoder is frozen and bone/soft-tissue vectors are extracted deterministically. Downstream models operate on these features rather than updating the encoder unless a separate fine-tuning experiment explicitly does so.
Baseline models
The general-image comparison is frozen DINOv3; the tabular comparison uses scanner-derived measurements. Age, sex and body mass index are included consistently in the relevant feature-based tasks. Numerical preprocessing and model selection are fitted within the specified splits. These are comparisons of complete feature-and-readout pipelines, not encoder architecture alone.
Downstream evaluation in HPP
Linear or regularized readouts predict continuous traits and binary diagnoses. Hyperparameters are selected within subject-disjoint training folds, and performance is repeated across ten random splits. A separate end-to-end fine-tuning sensitivity study updates both image models under a shared protocol. Organ-system labels aggregate diagnoses using the supplementary mapping, which differs from testing each individual disease.
UKBB cross-sectional disease classification
The external encoder remains frozen, but the external cohort supplies labels for fitting and validating a new logistic-regression readout. External transfer of an encoder therefore does not mean a complete classifier trained only in the original cohort. The four comparison arms use the same subject split and applicable covariates.
Longitudinal incident-disease risk
Participants with the endpoint already recorded at baseline are excluded. Penalized Cox models use scan features and covariates, with tuning and repeated holdout evaluation. A concordance index measures ordering of event risk; censoring handles follow-up that ends before an event is observed. Top-quartile case capture provides another view of ranking, without establishing a treatment threshold or net clinical benefit.
Biological age
Out-of-fold age predictions avoid predicting a person using a regressor fitted on that person. Removing linear and quadratic dependence on chronological age constructs the gap; near-zero age correlation is consequently partly by design. Subsequent analyses compare phenotypes, disease prevalence, mortality and medication-associated change, with their own models and adjustments. Residual confounding and small medication groups remain relevant.
Statistical analysis
Regression uses correlation and classification uses AUROC. Multiple-testing correction limits the expected fraction of false discoveries among declared discoveries under its assumptions; an adjusted significance threshold does not measure effect size or practical importance. Repeated splits reuse many participants, so the reported variation across splits is different from collecting ten independent cohorts.
Equation


Clustering
The pipeline standardizes features, reduces them with principal components, creates a two-dimensional UMAP map and fits two clusters there. Outlier filtering precedes one-to-one matching on age and body mass index. Molecular comparisons use platform-specific subsets with different sample sizes. These choices make the cluster analysis exploratory and dependent on the representation, projection and matching procedure.
Genome-wide association of embedding space
The study tests the leading 20 image-vector components against genetic variants after quality control and adjustment for demographic and genetic ancestry components. Nearby correlated variants are grouped into loci. Heritability estimates concern variance explained by measured common variants under the analysis model. Comparing principal components of image features with individual tabular traits is informative, but those are different phenotype constructions.
Data availability
Research access to the Human Phenotype Project data requires its stated institutional application process. The paper does not make individual medical scans freely downloadable.
Code availability
The authors link their implementation and released result tables. The exact pinned manuscript, rather than a changing repository, is the authority for this edition’s reported protocol.
Acknowledgments
Funding and acknowledgments are retained in the source text. They do not form part of the method or evaluation evidence.
Additional information
The original manuscript identifies the corresponding author. The reading edition preserves the paper’s provenance and complete bibliography.
Figure S1 — Architecture and evaluation setup
The diagram separates crop-based, label-free pretraining from supervised downstream readouts. SIGReg regularizes projected training coordinates; the final frozen encoder supplies the scan features.

Figure S2 — Cohort selection
Follow the participant counts through image availability and completeness filters. The external pretraining exclusion and task-specific analytic sample sizes answer different questions.

Figure S3 — Recovering conventional scan measurements
Reconstructing familiar scanner readouts checks whether the features retain known information. Bone-density classification also uses regional scans; it is not the same input protocol as every full-body task.

Figure S4 — Organ-system disease prediction
Grouping diagnoses tests broader outcome categories. Categories contain different numbers and mixtures of diseases, so their scores cannot be interpreted as equally difficult tasks.

Figure S5 — Fine-tuning sensitivity
Updating the encoders under a shared supervised protocol broadly preserves the comparison on eight selected targets. This sensitivity experiment is distinct from the main frozen-feature results.
![Figure S5 | Supervised fine-tuning does not change model performance rankings. Frozen linear probing (open markers) versus full end-to-end supervised fine-tuning (filled markers) for LeDXA and DINOv3, using embeddings alone (no age/sex/BMI covariates). Data represent mean ± standard error of the mean (s.e.m.) across 5 random seeds. a, Prediction of three continuous physiological biomarkers (chronological age, creatinine, and hemoglobin), evaluated by Pearson correlation (r). b, Classification performance (area under the receiver operating characteristic curve [AUROC]) for five prevalent conditions in the Human Phenotype Project (HPP) cohort: metabolic dysfunction-associated steatotic liver disease (MASLD), sleep apnea, osteoarthritis, anemia, and anxiety. Across all eight evaluated targets, supervised fine- tuning did not reverse the performance ranking of LeDXA relative to DINOv3 established under frozen probing.](https://storage.literature.yihao.one/media/sha256/62/62f4e0302f896f965f06397dc6fb0ac4296e5c9e1ae4f6b2c7b55f1a7b1af379.webp)
Figure S6 — Age-gap detrending
The residual has little remaining correlation with chronological age after detrending. This is an intended property of the score construction, rather than independent proof that it measures aging correctly.

Figure S7 — Internal replication of age-gap patterns
The internal cohort broadly reproduces selected directions of age-gap associations. This supports consistency of those associations, without converting them into causal effects.

Figure S8 — Disease prevalence across age-gap quartiles
The bars compare disease prevalence among people with younger- and older-looking scans. These are cross-sectional associations and differ from predicting newly occurring disease during follow-up.

Table S1 — Cohort characteristics
The cohorts differ in age, country, sex distribution and body composition. Reported sample sizes depend on the completeness filter used for this table; they need not equal all available scans.
Table S2 — Internal disease classification
The full table includes wins, ties and losses across diseases. Read each AUROC alongside its case count, variation and covariate-adjusted baseline, rather than using selected examples alone.
Table S3 — External disease classification
These results fit task readouts in UK Biobank using the frozen encoder. External feature transfer and external validation of a fully fixed diagnostic system are different evaluation designs.
Table S4 — Organ-system grouping
This mapping defines the aggregated labels used in the system-level analysis. A participant is positive when at least one included diagnosis is present.
Table S5 — Incident-disease risk
The complete concordance-index table shows the endpoint-specific picture, including outcomes without a LeDXA advantage. More accurate ranking does not by itself establish clinical utility.
Table S6 — Genetic association hits
Variants and nearest-gene annotations help inspect known and representation-specific signals. Nearby-gene assignment and statistical association do not identify a proven causal gene.
Table S7 — Medication-associated age-gap change
The table retains all tested medication groups, including nonsignificant results and small paired samples. Before/after changes are observational and can reflect changes beyond the medication itself.
Table S8 — Matched cluster phenotypes
These feature comparisons describe the two female clusters after matching on age and body mass index. Other differences can remain; matching is not random assignment.
Table S9 — Cluster molecular measurements
Different molecular platforms cover different subsets. The table includes significant and suggestive comparisons, so the adjusted significance values matter when distinguishing supported findings.
Table S10 — Scanner feature dictionary
The baseline combines body composition, bone density, geometry and related scanner quantities. The dictionary makes clear that this is a substantial measured-feature baseline, rather than only one bone-density number.