The ReiefF algorithm, model sensitivity and specificity and the receiver operated characteristic (ROC) curve.

(A) Running the ReliefF algorithm yields an ordered list of features that best separate the classes. Here the best 600 features are plotted against their utility in separation. The first 50 feature values are shown in detail (Inset). Note that the major inflection point occurs at approximately feature 125. (B) Ten-fold cross-validation sensitivity (mustard) and specificity (blue) as a function of the number of first N features used in the model. Of all parameter choices (see Table 1), the two best performed SVM, RF and LR models are shown in the first, second and third rows respectively. Note that the SVM outperforms the RF and LR models, and note that incorporating additional features beyond 125 does not improve performance, an observation that is independent of the choice of algorithm. (C) Leave-one-out cross validation ROC curve utilising the Shannon4 wavelet decomposition and the SVM. The SVM outputs a probability of diagnosis with respect to the classes. The upper left of the curve reveals two important points dependent on the decision probability cutoff: a sensitivity of 0.97 and a specificity of 0.89; and a sensitivity of 0.93 and a specificity of 0.93. Both points are equally accurate, but the latter is closer to the point (0,1).