Combined multivariate matching pursuit and support vector machine: A way forward to classify single-sweep evoked potentials?

Evoked brain potentials averaged over multiple sweeps provide a valuable objective measure of abnormal pain processing due to sensitization of the central nervous system. However, the average procedure cancel out important information regarding phase resetting and non-phase locked oscillations. Hence, assessment of the pain processing could be optimized by analyzing single-sweeps. To develop improved methods to assess single-sweeps, we applied a new approach in one healthy volunteer participating in a placebo controlled study of widespread hyperalgesia induced by perfusion of acid and capsaicin in the esophagus. The evoked potentials were recorded during electrical stimulations in the rectosigmoid colon. Features from the single-sweeps were extracted by a multivariate matching pursuit algorithm with Gabor atoms, and features were discriminated by a support vector machine with a linear kernel. The classification performance for the optimal number of atoms was 95% when discriminating the sensitization response from the placebo response, which was above change level compared to the performance when discriminating the two baseline responses (P <; 0.001). The discriminative capacity was increased power in the delta, theta, and alpha frequency bands. This result corresponds to previous characteristics seen in chronic pain patients who exhibit central sensitization. The new approach to classify single-sweeps on a single subject basis might in the future prove to be a useful tool in assessing mechanisms in central sensitization, and could be applied to improve enriched enrollment of study subjects in clinical trial units.