Model comparison for automatic characterization and classification of average ERPs using visual oddball paradigm.

OBJECTIVE To determine whether automated classifiers can be used for correctly identifying target categorization responses from averaged event-related potentials (ERPs) along with identifying appropriate features and classification models for computer-assisted investigation of attentional processes. METHODS ERPs were recorded during a target categorization task. Automated classification of average target ERPs versus average non-target ERPs was performed by extracting different combinations of features from the P300 and N200 components, which were used to train six classifiers: Euclidean classifier (EC), Mahalanobis discriminant (MD), quadratic classifier (QC), Fisher linear discriminant (FLD), multi-layer perceptron neural network (MLP) and support vector machine (SVM). RESULTS The best classification performance (accuracy: 91-92%; sensitivity: 85-86%; specificity: 95-99%) was provided by QC, MLP, SVM on feature vectors extracted from P300 recorded at multiple sites. In general, non-linear and non-parametric classifiers (QC, MLP, SVM) performed better than linear classifiers (EC, MD, FLD). The N200 did not explain variance beyond that of P300 recorded at multiple sites. CONCLUSIONS The results suggest that automatic characterization and classification of average target and non-target ERPs is feasible. Features of P300 recorded at multiple sites used to train non-linear classifiers are recommended for optimal classification performance. SIGNIFICANCE Automatic characterization of target ERPs can provide an objective approach for detecting and diagnosing abnormalities and evaluating interventions for clinical populations, paving the way for future real-time monitoring of attentional processes.