W. Pedrycz & S.-M. Chen (Eds.): Time Series Analysis, Model. & Applications, ISRL 47, pp. 249–278. DOI: 10.1007/978-3-642-33439-9_12 © Springer-Verlag Berlin Heidelberg 2013 Chapter 12 Channel and Class Dependent Time-Series Embedding Using Partial Mutual Information Improves Sensorimotor Rhythm Based Brain-Computer Interfaces Damien Coyle * Abstract. Mutual information has been found to be a suitable measure of depen- dence among variables for input variable selection. For time-series prediction mu- tual information can quantify the average amount of information contained in the lagged measurements of a time series. Information quantities can be used for select- ing the optimal time lag, τ, and embedding dimension, Δ, to optimize prediction ac- curacy. Times series modeling and prediction through traditional and computation- al intelligence techniques such as fuzzy and recurrent neural networks (FNNs and RNNs) have been promoted for EEG preprocessing and feature extraction to max- imize signal separability to improve the performance of brain-computer interface (BCI) systems. This work shows that spatially disparate EEG channels have differ- ent optimal time embedding parameters which change and evolve depending on the class of motor imagery (movement imagination) being processed. To determine the optimal time embedding for each EEG channel (time-series) for each class an ap- proach based on the estimation of partial mutual information (PMI) is employed. The PMI selected embedding parameters are used to embed the time series for each channel and class before self-organizing fuzzy neural network (SOFNN) based predictors are specialization to predict channel and class specific data in a predic- tion based signal processing framework, referred to as neural-time-series- prediction-preprocessing (NTSPP). The results of eighteen subjects show that subject-, channel- and class-specific optimal time embedding parameter selection using PMI improves the NTSPP framework, increasing time-series separability. The chapter also shows how a range of traditional signal processing tools can be combined with multiple computational intelligence based approaches including the SOFNN and practical swarm optimization (PSO) to develop a more autonomous parameter optimization setup and ultimately a novel and more accurate BCI. Damien Coyle Intelligent Systems Research Centre, University of Ulster, Derry, BT48 7JL, UK