In silico augmentation strategies for enhanced machine learning performance in fracture recognition

Abstract

This study presents a machine learning framework for fracture risk prediction and in silico validation of synthetic biomedical data. A retrospective dataset comprising 169 patient records with clinically relevant variables, including age, sex, weight, height, medication status, and bone mineral density (BMD), was analyzed. Multiple classification models, including Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine, and ensemble voting classifiers, were evaluated using 5-fold stratified cross-validation. Synthetic data fidelity was assessed through statistical distribution alignment, correlation preservation, and predictive transferability between real and synthetic domains. Among the evaluated models, the Voting Hard ensemble achieved the highest classification performance with an accuracy of 85.8% and F1-score of 0.822, while Logistic Regression demonstrated the highest discriminative capability (AUC = 0.88). Synthetic data showed strong agreement with real data in marginal feature distributions but weaker preservation of inter-feature correlations. The findings demonstrate the potential of ensemble machine learning methods for fracture risk prediction while highlighting the importance of rigorous validation when utilizing synthetic biomedical datasets. This framework provides a foundation for future development of privacy-preserving and clinically relevant synthetic data applications in biomedical machine learning.