Comparison of four prediction models to discriminate benign from malignant vertebral compression fractures according to MRI feature analysis

OBJECTIVE. The purpose of this study was to identify a set of significant predictors, predominantly based on MRI features and limited demographic data, for differentiating benign from malignant vertebral compression fractures (VCFs) and to compare the diagnostic value of four different types of prediction models. MATERIALS AND METHODS. The cohort included 101 subjects (mean age, 64 years; range, 22-89 years; 39.6% were men) with 128 VCFs, 84 (65.6%) of which were proven to be malignant by biopsy or follow-up imaging. A set of 36 predictors (34 MRI features, age, and sex) was acquired for each VCF retrospectively. RESULTS. Univariate analysis and initial variable screening selected 14 moderately significant predictors. All four model types performed similarly in overall discrimination power. The median area under the curve for each model type was 0.872 for logistic regression, 0.781 for classification tree, 0.760 for support vector machine, and 0.730 for neural network, but no significant difference was found between any two model types by Student t test. Five predictors for the logistic regression model were statistically significant (p < 0.01). Focal paraspinal mass and depositlike appearance of pedicle involvement were positively correlated with VCF malignancy. Age, other compression deformities without bone marrow edema, and intravertebral fluid collection or fluid signal were negatively correlated with VCF malignancy. CONCLUSION. It is possible to estimate the malignancy risk of VCF by using a small number of MRI features and patient age. The diagnostic performance of models selected by logistic regression, support vector machine, neural network, and classification tree was similar.