Human computer interface with online brute force feature selection

Abstract

Biosignal based human computer interfaces (HCI) become increasingly relevant for assistive, rehabilitation, or enter-tainment purposes. Practical systems need a continuous recalibration of feature and classifier settings from beginning of the session. Our long-term objective is the development of a versatile pattern recognition subsystem for HCI applications with automatic calibration for different kinds of biosignals. In this study the feasibility of an online brute force feature selection was examined and initially tested with electroencephalogram (EEG) based brain computer interface (BCI) data. Five of six subjects could control a synchronous two-class driving game within 80 trials. The algorithm found suitable features and linear classifiers in parallel to the running experiment. Applied to the BCI Competition IV data set 2b the method reaches appropriate accuracies. The presented strategy is relevant for systems on demand, because it makes it possible to avoid long lasting offline calibration procedures and to provide an environment with a single start button. Some exploratory tests revealed evidence for working with other biosignals as well. © 2012 by Walter de Gruyter Berlin Boston.