REVOLUTIONIZING FRACTURE RISK PREDICTION: MACHINE LEARNING AND GENETIC INSIGHTS TO ADDRESS DISPARITIES

Osteoporotic fractures occur at alarming rates globally, with over 8.9 million fractures annually, disproportionately affecting underserved populations. Despite its widespread use, the Fracture Risk Assessment Tool (FRAX) fails to adequately account for genetic variability and racial differences, leading to significant inaccuracies in fracture risk prediction. These limitations leave African American and Asian women, among others, at heightened risk of undetected fractures, exacerbating healthcare disparities. Addressing this challenge requires a paradigm shift—integrating genetic insights with advanced machine learning (ML) techniques to enhance precision, equity, and outcomes in fracture care. To identify the limitations of FRAX and develop an innovative, equity-focused predictive model integrating GRS with ML techniques. Methods. This study analyzed data from 25,772 postmenopausal women in the Women's Health Initiative (WHI) to evaluate FRAX performance across varied racial and genetic risk profiles. GRS was derived using genome-wide association studies (GWAS) findings. The WHI dataset was randomly divided into training (80%) and testing (20%) subsets. Four ML models—Support Vector Machine, Random Forest, XGBoost, and Artificial Neural Network—were developed to predict binary fracture outcomes and 10-year fracture risk. Bayesian optimization, a robust hyperparameter tuning method, was employed to enhance model performance, outperforming traditional grid search techniques. Additionally, death as a competing risk was accounted for in time-to-event analyses. Model performance was evaluated using accuracy, area under the curve (AUC), and net reclassification improvement (NRI). Portability was assessed across diverse racial groups, including African American and Asian populations. FRAX significantly overestimated fracture risk in low-GRS groups (predicted-to-observed ratio [POR]: 1.61; 95% CI: 1.45–1.79) and racial minorities, including Asian (POR: 3.50; 95% CI: 2.48–4.81) and African American women (POR: 2.59; 95% CI: 2.33–2.87). The XGBoost model, optimized with Bayesian methods, outperformed FRAX (accuracy: 91.2%; AUC: 0.739; C-index: 0.795) and achieved a categorical NRI of 22.6% ( P = 0.004). The model demonstrated a mean dynamic AUC of 0.799 for time-to-event predictions, surpassing FRAX in predicting 10-year fracture risk. Robust portability across racial groups, including African American and Asian populations, underscored its potential to mitigate disparities and improve clinical utility. This study presents a transformative framework combining GRS and ML to redefine fracture risk prediction. By addressing long-standing disparities and improving precision, this ML model paves the way for equitable, personalized fracture care, setting a new benchmark in osteoporosis management. Its integration into clinical practice and policymaking could revolutionize fracture prevention and improve health outcomes globally.