Deep learning predicts all-cause mortality from longitudinal total-body DXA imaging.

Area of Science:

• Gerontology
• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Biomarkers like BMI predict mortality, but longitudinal changes and advanced imaging are underutilized.
• Existing mortality prediction models often lack the granularity of full-body imaging data.
• This study explores novel approaches using DXA imaging for enhanced mortality risk assessment.

Purpose of the Study:

• To test if deep learning features from total-body DXA scans predict all-cause mortality.
• To evaluate if sequential DXA scans and recurrent neural networks improve mortality prediction over single observations.
• To assess model performance with and without traditional clinical risk factors.

Main Methods:

• Developed and tested multiple deep neural network architectures.
• Utilized data from the Health, Aging, and Body Composition Study (over 15,000 scans, 3000+ participants).
• Employed explainable AI techniques to interpret model predictions and feature importance.

Main Results:

• Longitudinal total-body DXA scans significantly predict all-cause mortality.
• The strongest deep learning model achieved an area under the ROC curve of 0.79 on a test set.
• Deep learning models incorporating longitudinal DXA data outperformed traditional models.

Conclusions:

• Deep learning effectively analyzes DXA medical imaging for mortality prediction in both cross-sectional and longitudinal contexts.
• Sequential DXA scan analysis with recurrent neural networks enhances mortality prediction accuracy.
• This research provides insights into healthy aging markers within a diverse population.