Computer-Aided Detection of Incidental Lumbar Spine Fractures from Routine Dual-Energy X-Ray Absorptiometry (DEXA) Studies Using a Support Vector Machine (SVM) Classifier

To assess whether application of a support vector machine learning algorithm to ancillary data obtained from posterior-anterior dual-energy X-ray absorptiometry (DEXA) studies could identify patients with lumbar spine (L1–L4) vertebral body fractures without additional DEXA imaging or radiation. Three hundred seven patients (199 without any fractures of the spine, and 108 patients with at least one fracture of the L1, L2, L3, or L4 vertebral bodies) who had DEXA studies were evaluated. Ancillary data from DEXA output was analyzed. The dataset was split into training (80%) and test (20%) datasets. Support vector machines (SVMs) with 10-fold cross-validation and different kernels were used to identify the best kernel based on the greatest area under the curve (AUC) and the best training vectors in the training dataset. The SVM with the best kernel was then applied to the test dataset to assess the accuracy of the SVM. Receiver operating characteristic (ROC) curves of the SVMs using different kernels in the test dataset were compared using DeLong’s test. The SVM classifier with the linear kernel had the greatest AUC in the training dataset (AUC = 0.9258). The AUC of the SVM classifier with the linear kernel in the test dataset was 0.8963. The SVM classifier with the linear kernel had an overall average accuracy of 91.8% in the test dataset. The sensitivity, specificity, positive predictive value, and negative predictive of the SVM classifier with the linear kernel to detect lumbar spine fractures were 81.8%, 97.4%, 94.7%, and 90.5%, respectively. The SVM classifier with the linear kernel ROC curve had a significantly better AUC than the SVM classifier with the cubic polynomial kernel ( P  = 0.034) for discriminating between patients with lumbar spine fractures and control patients, but not significantly different from the SVM classifier with a radial basis function (RBF) kernel ( P  = 0.317) or the SVM classifier with a sigmoid kernel ( P  = 0.729). All fractures identified by the SVM classifiers were not prospectively identified by the radiologist. SVM analysis of ancillary data obtained from routine DEXA studies can identify lumbar spine fractures without the use of vertebral fracture assessment (VFA) DEXA imaging or radiation, and identify fractures missed by radiologists.