A comprehensive analysis and performance evaluation for osteoporosis prediction models.

Area of Science:

• Medical data analysis and computational health.
• Application of artificial intelligence in healthcare diagnostics.

Background:

• Medical data analysis offers transformative potential for healthcare.
• Leveraging research data enhances clinical decision-making and patient outcomes.

Purpose of the Study:

• To predict osteoporosis onset using deep learning techniques.
• To evaluate feature selection methods (mutual information and recursive feature elimination) for deep neural network models.

Main Methods:

• Utilized the NHANES 2017-2020 dataset, preprocessed into SpineOsteo and FemurOsteo datasets.
• Applied sequential deep neural networks, convolutional neural networks (CNN), and recurrent neural networks.
• Employed mutual information (MI) and recursive feature elimination (RFE) for feature selection.

Main Results:

• Mutual information (MI) outperformed recursive feature elimination (RFE) in accuracy.
• The MI-selected CNN model achieved 99.15% accuracy for SpineOsteo and 99.94% for FemurOsteo.
• Identified significant predictors: family medical history, patient fractures, parental hip fractures, and regular use of prednisone or cortisone.

Conclusions:

• Deep learning, particularly CNN with MI feature selection, demonstrates high efficacy in osteoporosis prediction from non-image medical data.
• The findings support enhanced diagnostic and prognostic models for healthcare providers.
• Highlights the importance of specific clinical and familial factors in osteoporosis risk assessment.