Skip to content

Shepherd Research Lab

    Deep Learning and Total Body DXA Scans (The TBDXA.I. Project)

    Overview

    Deep Learning trains neural networks on large datasets, including images, to generate algorithms that predict endpoints. Total Body DXA scans contain digital information about body composition (fat, bone, lean tissue) and its distribution and body shape. Each of these characteristics (bone mass, fat mass, and lean mass) has been shown to be individually predictive of clinical outcomes associated with aging.

    In this project, we attempt to use deep learning methods on total body DXA scans to extract more information than was previously done, and thus, providing more accurate predictions of clinical outcomes, including cardiovascular disease (CVD), CVD death, overall mortality, cancer, cancer death, hip fracture, physical disability, incident insulin-resistant diabetes, and severity of insulin resistance.

    We will use a novel new approach called self-supervised learning to extract features from the DXA whole body images collected in the Health, Aging and Body Composition (Health ABC) Study. This dataset includes more than 3,000 participants, along with their follow-up examinations in years 3, 6, and 10. This learning process does not use the aging outcomes, just the “unlabeled” images themselves. It will create a penultimate set of features found in the DXA images that can be used in the neural network or outside it to predict the outcomes of interests.

    Objective/Aims

    The focus of this project is to apply deep learning methods to baseline and follow-up total Body DXA images to predict outcomes associated with age-related changes in body composition.

    Our specific aims are:

    1. Generate predictive algorithms for CVD, CVD death, overall mortality, major disease-free mortality, cancer, cancer death, hip fracture, physical disability, incident insulin-resistant diabetes, and hemoglobin A1c, and fasting insulin and glucose levels by applying deep learning methods to baseline and follow-up Total Body DXA images.
    2. Generate predictive algorithms for physical performance including gait speed and long-distance corridor walk speed and levels of inflammation (eg. IL-6 concentration).
    3. Explore features of TBSXA images that account for the predictive accuracy of the algorithms by using saliency mapping.

    Research Team

    Steve Cummings, MD, FAPC

    Principal Investigator

    SF Coordinating Center

    John Shepherd, PhD

    Principal Investigator

    UH Cancer Center

    Peter Sadowski profile photo

    Peter Sadowski, PhD

    Co-Investigator

    UH Manoa Information and Computer Sciences

    Warren Browner, MD, MPH

    CEO

    California Pacific Medical Center

    Eleanor Simonsick, PhD

    Epidemiologist

    National Institute on Aging – NIH

    Yannik Glaser

    Graduate Student

    UH Manoa Information and Computer Sciences

    Lily Liu

    Statistician

    California Pacific Medical Center

    Funding Source(s)

    Publications & Presentations

    1.
    Glaser Y, Shepherd J, Leong L, Wolfgruber T, Lui LY, Sadowski P, et al. Deep learning predicts all-cause mortality from longitudinal total-body DXA imaging. Commun Med [Internet]. 2022 Aug 16;2. Cite Download
    1.
    Glaser Y, Sadowski P, Wolfgruber T, Lui LY, Cummings S, Shepherd J. Hip Fracture Risk Modelling Using DXA and Artificial Intelligence. Poster presented at: American Society for Bone Mineral and Research Annual Meeting; 2020 Sep 11; Virtual. Cite Download