Developing and comparing deep learning and machine learning algorithms for osteoporosis risk prediction.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Gerontology

Background:

• Osteoporosis, defined by low bone mineral density (BMD), presents a significant public health challenge.
• Existing regression and machine learning (ML) models for osteoporosis risk prediction exhibit limited accuracy in clinical settings.
• Deep learning (DL) approaches, like deep neural networks (DNNs), offer potential for improved prediction by uncovering complex data interactions.

Purpose of the Study:

• To evaluate the performance of a novel DNN framework in predicting osteoporosis risk.
• To compare the predictive accuracy of the DNN model against conventional ML algorithms and a traditional regression model.
• To identify key predictive variables for osteoporosis risk using feature importance analysis.

Main Methods:

• A novel DNN framework was developed using hip BMD and clinical data from 8,134 subjects over 40 years old from the Louisiana Osteoporosis Study (LOS).
• The DNN model's performance was compared against Random Forest (RF), Artificial Neural Network (ANN), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and Osteoporosis Self-Assessment Tool (OST).
• Performance was evaluated using Area Under the Receiver Operating Curve (AUC) and accuracy metrics.

Main Results:

• The DNN approach achieved the highest predictive performance (AUC = 0.848) in classifying osteoporosis risk.
• Key predictive variables identified by DNN include weight, age, gender, grip strength, and height.
• The DNN model maintained high performance (AUC = 0.846) even with reduced feature sets (top 10 variables) and sample sizes (50% reduction).

Conclusions:

• A novel DNN model demonstrates effectiveness for the early diagnosis and intervention of osteoporosis in aging populations.
• DNNs offer a promising advancement over traditional methods for osteoporosis risk prediction.
• The model's robustness with reduced data highlights its potential for practical clinical application.