Artificial intelligence-based diabetes risk prediction from longitudinal DXA bone measurements.

Area of Science:

• Endocrinology and Metabolism
• Medical Imaging
• Gerontology

Background:

• Diabetes mellitus (DM) is a global health concern with severe complications.
• Predicting diabetes onset using longitudinal multi-modal data is underexplored.
• Dual-energy X-ray absorptiometry (DXA) measures bone composition, potentially offering insights into metabolic health.

Purpose of the Study:

• To investigate DXA-derived bone composition measures as predictors of diabetes onset in Qatari adults.
• To explore risk factors associated with diabetes development using longitudinal multi-modal data.
• To compare conventional and deep learning models for diabetes risk prediction.

Main Methods:

• A retrospective case-control study of 1,382 Qatari adults.
• Data augmentation using SMOTE and SMOTEENN to handle class imbalance.
• Analysis of bone mineral density (BMD) and bone mineral content (BMC) using ANOVA, SHAP, and probabilistic methods.
• Comparison of shallow and deep learning models for diabetes prediction.

Main Results:

• Diabetic participants exhibited higher BMD and BMC in multiple regions compared to controls.
• Abnormal glucose metabolism correlated with increased BMD and BMC but lower Z-scores.
• Shallow learning models achieved superior prediction accuracy (91.08%), AUROC (96%), and recall (91%) over deep learning models.
• Diabetes prediction accuracy improved for older age groups and was higher in males.

Conclusions:

• DXA-derived bone composition measures are promising predictors for early diabetes detection.
• This minimally invasive approach using DXA scans offers potential for rapid risk assessment.
• Understanding bone health markers can contribute to proactive diabetes management strategies.