Multivariate pattern analysis of evoked brain potentials by temporal matching pursuit and support vector machine

Abstract

Background/aims Electroencephalography (EEG) recorded as evoked brain potentials (EPs) reflects the cortical processing to an external event. This approach is often used to study the altered response to acute pain in chronic pain patients compared to healthy volunteers. However, discrimination of the responses from the study populations is a non-trivial task, which calls for improved objective methods. Methods To develop and validate a new methodology, we analyzed data from 16 type-1 diabetes mellitus patients and 15 age and gender matched volunteers, by means of brain activity recorded from 62 EEG channels. The EEG signals were recorded as EPs elicited by painful electrical stimulations in the oesophagus with an intensity corresponding to the individual pain detection threshold. The EPs from all channels and subjects were decomposed simultaneously by a temporal matching pursuit (TMP) algorithm with Gabor atoms. Results Amplitude and phase features were classified by a support vector machine (SVM) to discriminate patients from healthy volunteers. A classification performance of 93.1% (P<0.001) was obtained when applying a majority voting scheme to the 3 best performing channels (FC4, C1, and C6) and including features from 2 atoms. The most discriminative features were determined by the slope coefficients from the SVM decision rule, which identified the biomarkers as delayed latency of the first atom (N2–P2 complex) and decreased amplitude of the second atom (NI–P1 complex). Conclusion The combination of TMP and SVM is a novel approach to classify two study populations, which may provide a new objective tool to identify biomarkers from various chronic pain populations.