Contributors: Thorsten Zander, FIRST FhG (Student) Guido Dornhege, FIRST FhG Benjamin Blankertz, FIRST FhG This approach works on AAR models, whose coefficients where calculated for 'C3' and 'C4' on each single trial [1]. A weight vector in time, which reflects the difference in ERD between the given classes, is determined by an optimization procedure [2]. For this purpose the time course of the amplitude of the mu-rythm was estimated by calculating the band power 9-11 Hz in the power spectrum obtained from the AAR models. This was performed for different choices of the AAR-model-parameters 'order' and 'update coefficient'. Some of the resulting features were combined to optimize the cross-validation performance of a linear classifier [3]. For calculating the confidence at time 't' of a test trial, the trained classifier is applied to the (incomplete) trial up to and including 't' and the output is weighted by the above given ERD measure, normalized by the intra-class variances (of the estimated classes) and sumed up. So this approach works in a causal way. The confidence and classification results are given from 2s to 8.85s. [1] The AAR models were calculated by a matlab implementation of A. Schloegel,TU Graz [2] Based on an algorithm by Michael Zibulevsky, Technion Israel [3] Feature Combination, Guido Dornhege FIRST FhG Thanks for helpful talks with Steven Lemm and Christin Schaefer both FIRST FhG