Osteoporotic vertebral fractures are a major cause of morbidity, disability, and mortality among the elderly. Traditional methods for fracture risk assessment, such as dual-energy X-ray absorptiometry (DXA), may not fully capture the complex factors contributing to fracture risk. This study aims to enhance vertebral fracture risk prediction by integrating radiomics features extracted from computed tomography (CT) scans with clinical data, utilizing advanced machine learning techniques.
We analyzed CT imaging data and clinical records from 124 patients, extracting a comprehensive set of radiomics features. The dataset included shape, texture, and intensity metrics from segmented vertebrae, alongside clinical variables such as age and DXA T-values. Feature selection was conducted using a Random Forest model, and the predictive performance of multiple machine learning models—Random Forest, Gradient Boosting, Support Vector Machines, and XGBoost—was evaluated. Outcomes included the number of fractures (N_Fx), mean fracture grade, and mean fracture shape. Incorporating radiomics features with clinical data significantly improved predictive accuracy across all outcomes. The XGBoost model demonstrated superior performance, achieving an R2 of 0.7620 for N_Fx prediction in the training set and 0.7291 in the validation set. Key radiomics features such as Dependence Entropy, Total Energy, and Surface Volume Ratio showed strong correlations with fracture outcomes. Notably, Dependence Entropy, which reflects the complexity of voxel intensity arrangements, was a critical predictor of fracture severity and number.
This study underscores the potential of radiomics as a valuable tool for enhancing fracture risk assessment beyond traditional clinical methods. The integration of radiomics features with clinical data provides a more nuanced understanding of vertebral bone health, facilitating more accurate risk stratification and personalized management in osteoporosis care. Future research should focus on standardizing radiomics methodologies and validating these findings across diverse populations.