where x,z are training vectors. The SVM architectures were trained and tested on a 1.6 GHz Celeron using the D2CSVM software found at http://www.ee.unimelb.edu.au/people/dlai. We used the area of the receiver operating characteristics (ROC) curve for single features to obtain a quantitative measure of individual feature separability. ROC areas were numerically approximated using the trapezoidal rule where larger values implied better linear feature separability.
Table 1: Statistical Analysis and ROC Values of Individual Features for the Symptomatic Group (OA) and the Control Group.