LeDXA: Using DXA Scans to Identify Markers of Age and Risk

The LeDXA model detects features in whole-body X-ray images that are associated with age and the risk of future disease

On August 3, researchers released a preprint describing the LeDXA model. They trained it on 11 540 DXA scans, an X-ray examination that measures bone density and body composition, and tested it in 47 400 UK Biobank participants.

A DXA scanner usually reduces each image to a few measurements: bone density, fat mass, and lean mass in different parts of the body. The image itself also contains spatial details, including the distribution of tissues and the structure of the skeleton.

LeDXA receives two maps from the same examination: the distributions of bone mineral content and soft tissue. The model learned directly from 11 540 images by comparing different regions of the same body and constructing a numerical representation of its structure. The authors then applied it to DXA scans from UK Biobank that the model had not previously seen.

In this dataset, LeDXA estimated chronological age with a mean error of 2,9 years. Over a median follow-up of 4,3 years, it outperformed conventional DXA measurements in predicting nine of 20 diseases first diagnosed after the scan. For hip osteoarthritis, the quarter of participants with the highest predicted risk included 66% of future cases, whereas conventional DXA measurements identified 41%.

The authors also calculated a “biological age gap” by comparing the age predicted from the scan with the expected value for the person’s chronological age. Among 39 310 UK Biobank participants, the risk of death from any cause over a median follow-up of 3,4 years was 45,3% higher in the quarter with the largest gap than in the quarter with the smallest gap, after accounting for age and sex.

In a discussion of biological age, Matt Kaeberlein suggested evaluating such a measure by its association with future health. In this study, the same numerical representation provided both an age estimate and predictions for some later diagnoses. The comparison with conventional DXA measurements indicates that the spatial structure of the image contains additional prognostic information for these tasks.