Monday, December 1 | 2:20 p.m.-2:30 p.m. | M6-SSMK03-6 | Room E450A
In this scientific session, an AI model will be presented that is designed to predict 10-year mortality and hip fractures on dual-energy x-ray absorptiometry (DEXA) scans.
Notably, the model requires no additional imaging, clinical data, or patient burden, and thereby could enhance the value of routine DEXA exams in everyday practice, according to the researchers, a team from the University of Hawaii’s Artificial Intelligence Precision Health Institute led by Yannik Glaser, PhD.
The group developed a self-supervised vision transformer model and trained it without using labels to learn shared imaging patterns across 85,461 DEXA scans culled from U.S. National Institutes of Health-supported studies. The dataset included 33,360 whole-body, 28,928 hip, and 23,173 spine images.
Glaser and colleagues then tested the model on two clinical prediction tasks. For predicting 10-year all-cause mortality, they used 17,218 whole-body scans that included 6,911 patient deaths. For predicting 10-year hip fractures, they used 4,199 hip scans that included 258 fractures. The model’s outputs were based solely on extracted imaging features, without access to demographics, bone mineral density, or laboratory values.
The model achieved an area under the receiver operating characteristic curve (AUROC) of 0.7 for predicting 10-year mortality and an AUROC of 0.74 for predicting 10-year hip fracture.
Ultimately, the study demonstrates that DEXA images alone, when processed with advanced deep learning, can reveal imaging features predictive of long-term clinical outcomes, according to the team.