Abstract

While artificial intelligence has revolutionized medical diagnostics, osteoporosis prediction remains constrained by the non-personalized nature of conventional DXA scans and traditional risk assessments. This study addresses these limitations by developing a comparative machine learning framework utilizing datasets from Kaggle and NHANES, each comprising approximately 2,000 patient records. The data were standardized to encompass multifaceted risk factors, including physiological metrics, family history, and lifestyle habits. Ten distinct architectures ranging from traditional classifiers to deep learning models (CNN and RNN) and the BERT transformer were evaluated using 10-fold cross-validation. To handle data imbalance in the NHANES set, the SMOTE technique was applied. Notably, the BERT model demonstrated superior efficacy on the Kaggle dataset, achieving 92% accuracy and an AUC of 0.92. Furthermore, integrating a Genetic Algorithm for parameter optimization significantly enhanced the RNN’s performance, elevating its AUC to 0.932 and matching the transformer’s accuracy. These results suggest that combining deep learning with evolutionary optimization offers a robust, personalized approach for early osteoporosis detection, potentially improving long-term bone health management and preventive strategies.