Evolutionary optimization of classifiers and features for single-trial EEG Discrimination

BACKGROUND: State-of-the-art signal processing methods are known to detect information in single-trial event-related EEG data, a crucial aspect in development of real-time applications such as brain computer interfaces. This paper investigates one such novel approach, evaluating how individual classifier and feature subset tailoring affects classification of single-trial EEG finger movements. The discrete wavelet transform was used to extract signal features that were classified using linear regression and non-linear neural network models, which were trained and architecturally optimized with evolutionary algorithms. The input feature subsets were also allowed to evolve, thus performing feature selection in a wrapper fashion. Filter approaches were implemented as well by limiting the degree of optimization. RESULTS: Using only 10 features and 100 patterns, the non-linear wrapper approach achieved the highest validation classification accuracy (subject mean 75%), closely followed by the linear wrapper method (73.5%). The optimal features differed much between subjects, yet some physiologically plausible patterns were observed. CONCLUSION: High degrees of classifier parameter, structure and feature subset tailoring on individual levels substantially increase single-trial EEG classification rates, an important consideration in areas where highly accurate detection rates are essential. Also, the presented method provides insight into the spatial characteristics of finger movement EEG patterns.