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

Standard clinical DXA scans are primarily used to measure bone mineral density (BMD) and body composition, but these metrics discard the rich spatial information contained within the raw images. This study investigates whether self-supervised learning (SSL) can extract latent representations from these scans that capture systemic health, biological aging, and disease risk beyond what is currently available in clinical reports.

The researchers developed LeDXA, a vision model based on a Joint-Embedding Predictive Architecture (JEPA). Unlike models that attempt to reconstruct pixels, LeDXA learns by predicting latent representations, which encourages the model to focus on meaningful anatomical structures rather than scanner noise. The model was trained from scratch on 11,540 unlabeled whole-body DXA scans from the Human Phenotype Project (HPP) and subsequently validated on 47,400 external scans from the UK Biobank (UKBB).

LeDXA demonstrated superior performance in predicting both prevalent and incident diseases compared to traditional tabular DXA metrics and general-purpose models like DINOv3. Notably, in longitudinal follow-ups, LeDXA significantly improved the prediction of incident hip and knee arthrosis and type 2 diabetes. For hip arthrosis, the model successfully identified 66% of incident cases in its highest-risk quartile, compared to only 41% using standard clinical measures. Furthermore, the model's internal representations were highly predictive of chronological age and tracked biological aging, with the oldest-appearing quartile showing a 45% higher mortality hazard. The study also confirmed that these learned representations are heritable and align with known genetic loci for body composition.

This work demonstrates that routine medical imaging contains a wealth of prognostic information that is currently ignored in clinical practice. By using self-supervised learning, researchers can unlock these "hidden" signals to improve early disease detection and risk stratification without requiring expensive new diagnostic tests or massive, labeled datasets. The success of LeDXA suggests that domain-specific foundation models can be highly effective even when trained on relatively modest amounts of data.