Fig. 4.

Evaluation of a 3-way classification support vector machine (SVM) trained to classify between EEGs from subjects with lumbar radiculopathy ( N = 20), aged and gender matched healthy controls (pre-SCS, N = 20), and subjects with chronic back pain who are candidates for spinal cord stimulation ( N = 17). A) The Receiver Operating Characteristic (ROC) of the SVM with respect to each of the 3 classes, cross validated using leave-one-out cross-validation. The Area Under Curve (AUC) with respect to the radiculopathy group was 0.828, the AUC with respect to the healthy group was 0.842, the AUC with respect to the chronic pain group was 0.959, and the classification accuracy was 71.9%. B) A confusion matrix describing the cross-validation predictions of the trained SVM. 41 out of 57 samples were classified correctly (green squares) and 16 out of 57 were classified incorrectly (red squares). C) The results of 500 iterations of a random shuffling procedure, whereby the labels of the EEGs were randomly reassigned, and the SVM was retrained and retested using the randomized labels. The upper plot shows the histogram of classification accuracy amongst the 500 shuffled iteration in blue, and the true accuracy observed in our classifier in black (75.4%, as stated above). The observed accuracy was greater than that of all 500 shuffled iterations, indicating the odds of observing an accuracy of 75.4% given the null hypothesis that the classifier predictions are no better than chance is P < 0.002. The bottom shows the same results, except for AUC with respect to each class. For the healthy and Pre-SCS class, the observed AUC was greater than that of all of 500 shuffled iterations, indicating that P < 0.002. For the radiculopathy class, the observed AUC was greater than that of 499 out of 500 shuffled iterations, indicating that P = 0.002.