Modelling future bone mineral density: Simplicity or complexity?

Area of Science:

• Gerontology
• Medical Imaging
• Artificial Intelligence

Background:

• Osteoporotic fractures pose a significant global health burden, impacting patient well-being and healthcare costs.
• Bone mineral density (BMD) assessment is crucial for identifying osteoporosis and fracture risk.
• Current clinical algorithms lack effective methods to incorporate BMD changes over time into fracture risk prediction.

Purpose of the Study:

• To compare a statistical method (ZBM) with a deep learning (DL) based method for predicting future bone mineral density (BMD).
• To evaluate the performance of these models in forecasting BMD using longitudinal DXA data.

Main Methods:

• Longitudinal DXA data from 2948 adults (aged 40-90) with at least two hip scans were analyzed.
• A ZBM model predicted future BMD using reference group data and the latest scan.
• A DL-based method incorporated historical DXA data, ZBM features, and multidimensional longitudinal variables.

Main Results:

• Deep learning models, especially Hybrid-DL, significantly outperformed ZBM models in women.
• ZBM-based models performed comparably or better than DL-based models in men.
• The study included 2652 females (90%) and 296 males (10%).

Conclusions:

• Both DL-based and statistical models can forecast future BMD using longitudinal clinical data.
• These predictive models may enhance clinical decision-making for osteoporosis assessment, including timing of repeat BMD testing.