Oral Presentation 1st Asia Pacific Herbert Fleisch Workshop 2025

Machine learning based estimation of the appendicular lean mass index from hip and spine DXA scans in the UK Biobank Imaging Study (#28)

Marion Mundt 1 , Afsah Saleem 1 , Abadi Gebre 1 , William D Leslie 2 , Nam Ki Hong 3 , Gustavo Duque 4 , John Schousboe 5 , Nicholas Harvey 6 , Joshua Lewis 1 , Marc Sim 1
  1. Edith Cowan University, Joondalup, WA, Australia
  2. University of Manitoba, Winnipeg, Manitoba, Canada
  3. Yonsei University, Seoul, South Korea
  4. McGill University , Montreal, Quebec, Canada
  5. University of Minnesota, Minneapolis, Minnesota, USA
  6. University of Southampton, Southampton, UK

Introduction and aim: Appendicular lean mass can be assessed on whole-body dual energy X-ray absorptiometry (DXA) scans and are used to define sarcopenia1. Hip and anterior-posterior (AP) spine dual X-ray absorptiometry (DXA) scans are obtained to assess bone mineral density in more than 710,000 Australians annually2. These scans also contain important “clues” or information about appendicular lean mass (ALM)3,4.  We therefore applied a machine learning algorithm to hip and AP spine DXA scans to determine whether we could use these routinely captured images to estimate ALM.

Methods: Hip and AP spine DXA scans (GE Healthcare Lunar iDXA) and clinical data (sex, age, height, weight, BMI) of 41,506 individuals (51.4% female, age 64.4±7.8 years, height 169.2±9.3cm, weight 75.5±15.1kg, BMI 25.9±4.5kg/m2) from the UK Biobank Imaging Study5 were included. We developed a custom three-branch machine learning model with hierarchical fusion gates6 to predict the ALM index (ALMI = ALM/height2) obtained from whole-body DXA scans. Models were trained using 5-fold cross validation with 60% scans used in training, 20% in validation and 20% in testing for a maximum of 50 epochs and optimised using mean-squared-error loss.

Results: The model achieved an averaged root-mean-squared-error of 0.475 and R2 value of 0.859. Applying established cut-off values for low ALM (women<5.5kg/m2, men<7.0kg/m2)1, model accuracy was 92.7% (sensitivity 56.5%, specificity 97.4%, positive predictive value 74.4%, negative predictive value 94.5%, AUROC 77.0%). Predicted prevalence of low muscle mass was 8.8% compared to the measured prevalence of 11.6%.

Conclusion: Our results highlight the feasibility and potential of applying a machine learning algorithm to hip and spine DXA to estimate low ALM with high specificity as part of routine osteoporosis screening. This overcomes the need for additional whole-body scans in those with low ALM and provides clinicians with new and important information on an individual’s risk of sarcopenia/osteosarcopenia.

  1. 1. Cruz-Jentoft AJ, Bahat G, Bauer J, Boirie Y, Bruyère O, Cederholm T, Cooper C, Landi F, Rolland Y, Sayer AA, Schneider SM, Sieber CC, Topinkova E, Vandewoude M, Visser M, Zamboni M; Writing Group for the European Working Group on Sarcopenia in Older People 2 (EWGSOP2), and the Extended Group for EWGSOP2. Sarcopenia: revised European consensus on definition and diagnosis. Age Ageing. 2019 Jan 1;48(1):16-31. doi: 10.1093/ageing/afy169. Erratum in: Age Ageing. 2019 Jul 1;48(4):601. doi: 10.1093/ageing/afz046.
  2. 2. Australian Government. Medicare Statistics. Accessed 17 Feb, 2025.
  3. 3. Rossum K, Alexiuk MR, Bohm C, Leslie WD, Tangri N. Development of an Algorithm to Predict Appendicular Lean Mass Index From Regional Spine and Hip DXA Scans. J Clin Densitom. 2025 Apr-Jun;28(2):101560. doi: 10.1016/j.jocd.2024.101560.
  4. 4. Thackeray M, Orford NR, Kotowicz MA, Mohebbi M, Pasco JA. Estimation of Whole-Body and Appendicular Lean Mass from Spine and Hip Dual Energy X-ray Absorptiometry: A Cross-Sectional Study. Calcif Tissue Int. 2022 Mar;110(3):341-348. doi: 10.1007/s00223-021-00922-4.
  5. 5. Littlejohns TJ, Holliday J, Gibson LM, Garratt S, Oesingmann N, Alfaro-Almagro F, Bell JD, Boultwood C, Collins R, Conroy MC, Crabtree N, Doherty N, Frangi AF, Harvey NC, Leeson P, Miller KL, Neubauer S, Petersen SE, Sellors J, Sheard S, Smith SM, Sudlow CLM, Matthews PM, Allen NE. The UK Biobank imaging enhancement of 100,000 participants: rationale, data collection, management and future directions. Nat Commun. 2020 May 26;11(1):2624. doi: 10.1038/s41467-020-15948-9.
  6. 6. Arevalo J, Solorio T, Montes-y-Gómez M, González FA. Gated Multimodal Units for Information Fusion. arXiv [Preprint]. 2017 Feb 7:arXiv:1702.01992. Available from: https://arxiv.org/abs/1702.01992.