A Staged Approach using Machine Learning and Uncertainty Quantification to Predict the Risk of Hip Fracture.

Area of Science:

• Geriatric Medicine
• Radiology
• Biomedical Engineering

Background:

• Hip fractures are a major health concern in aging populations, often resulting from falls and leading to significant morbidity and mortality.
• Current methods for predicting hip fracture risk require improvement, especially considering factors like compromised bone quality.

Purpose of the Study:

• To develop and evaluate a novel staged model for predicting hip fracture risk in middle-aged and older adults.
• To improve predictive performance by integrating advanced imaging features with clinical data.

Main Methods:

• A staged machine learning model was developed using two ensemble models: one with clinical variables and another incorporating DXA imaging features.
• Convolutional Neural Networks (CNNs) were used to extract features from hip DXA images.
• Uncertainty quantification from the clinical-only model guided the decision to include DXA features.

Main Results:

• The model incorporating DXA features (Ensemble 2) achieved the highest performance with an AUC of 0.9541.
• The staged model demonstrated strong performance (AUC 0.8486) and suggested that 54.49% of patients may not require DXA scanning.
• The staged model significantly outperformed the model using only clinical variables (AUC 0.5549).

Conclusions:

• The proposed staged model offers a robust and cost-effective method for hip fracture risk prediction.
• This approach can accurately identify at-risk individuals while reducing the need for potentially unnecessary DXA scans, saving costs and radiation exposure.
• The model provides a holistic view of patient health, guiding targeted interventions for hip fracture prevention.