Machine Learning Principles Can Improve Hip Fracture Prediction.

Area of Science:

• Gerontology
• Biomedical Informatics
• Orthopedics

Background:

• Hip fractures pose a significant health burden, particularly in aging populations.
• Accurate prediction of hip fractures is crucial for timely intervention and prevention strategies.
• Existing prediction models often lack sufficient discrimination and calibration.

Purpose of the Study:

• To apply machine learning (ML) principles for predicting hip fractures.
• To estimate the importance of various predictors in hip fracture risk assessment.
• To compare the performance of ML models against traditional statistical methods.

Main Methods:

• Utilized Dual-energy X-ray absorptiometry (DXA) data from Danish regions (1996-2006) combined with national patient data.
• Developed 24 statistical models using k-5, 5-repeat cross-validation on 75% of data, validated on the remaining 25%.
• Employed ensemble ML models including bootstrap aggregated flexible discriminant analysis (bagFDA) and eXtreme Gradient Boosting (xgbTree).

Main Results:

• For women, bagFDA achieved a test Area Under the Curve (AUC) of 0.92, with an 11-predictor subset yielding AUC 0.91.
• For men, xgbTree achieved a test AUC of 0.89, with a 10-predictor subset yielding AUC 0.86.
• Machine learning models demonstrated improved hip fracture prediction and predictor importance estimation compared to logistic regression.

Conclusions:

• Machine learning, particularly ensemble methods, significantly enhances hip fracture prediction accuracy and calibration.
• Specific predictor subsets, including bone mineral density (BMD) and clinical factors, are vital for accurate risk assessment.
• Further research with international data and longer follow-up can refine ML models for broader clinical application.