Classification of Adaptive Autoregressive Models at Different Sampling Rates in a Motor Imagery-Based BCI Martin Billinger1, Clemens Brunner1 and Christa Neuper1,2 1 Institute for Knowledge Discovery, BCI Lab, Graz University of Technology, Austria 2 Department of Psychology, University of Graz, Austria Introduction Results 1 Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 Subject 6 Subject 7 Subject 8 Subject 9 0.9 0.8 0.7 0.6 Classification Accuracy (0.9-quantile) Autoregressive (AR) models have been employed for feature extraction in brain–computer interfaces (BCIs) using either AR coefficients directly [1] or the estimated spectrum [2]. From the perspective of spectral analysis, a model of order p can resolve p/2 frequency peaks. With sampling rate f s, the frequency resolution of an AR spectrum depends on the time window of length p/fs spanned by the AR model [2]. To increase the length of the time window, the model order can be increased, which may lead to overfitting caused by increased model complexity. Alternatively, reduction of fs does not increase model complexity, but limits bandwidth. This work attempts to provide an answer to the question: Can we improve AR–based classification of motor imagery by reducing the sample rate? 0.5 1 0.9 0.8 0.7 0.6 0.5 1 0.9 0.8 0.7 Methods 0.6 0.5 0 9 subjects Signal Processing A 160 trials AAR Resampling 2 classes B Bandpower 100 150 200 250 0 50 100 150 200 Sampling Rate [Hz] 250 0 50 100 150 200 250 Classification 10 x 10 running cross– classifier validation Bandpower 10-12, 16-24 Hz AR model order 6 Bandpower 10-12 Hz Random classification limit Figure 2: Individual results. The effect of sampling rate on classification accuracy varies considerably between subjects. LDA Figure 1: Block diagram of the offline system. For classification either AR or Bandpower features were utilized. Data: BCI Competition IV, data set 2b [3]. • 9 trained subjects • Motor imagery with feedback • 2 classes: left vs. right hand • 80 randomized trials per class Recording: • 3 bipolar EEG channels (C3, Cz, C4) • 250 Hz sampling rate Signal Processing: • Resampling to 5–245 Hz in 5 Hz steps • A: Adaptive AR Model ◦ Model order p = 6 ◦ Update coefficient U C = 10−5 • B: Bandpower ◦ 10–12 Hz ◦ 16–24 Hz Classification: • Linear Discriminant Analysis (LDA) • Running Classifier procedure finds optimal training time • 10×10 cross–validation • Subjects 2 and 3 did not perform above chance level. • Bandpower does not depend on sampling rate • Prominent rise in accuracy for subject 6 below 125 Hz mean accuracy 0.8 mean accuracy (250Hz) 0.78 95% confidence limit ACC Data 50 0.76 0.74 0.72 0.7 50 100 150 Sampling Rate [Hz] 200 250 Figure 3: Mean classification accuracy of all subjects. The dashed line shows the 95 % confidence limit of classification accuracy compared to 250 Hz sampling rate. • t–statistics for improvement of classification accuracy • significant improvement at 75 and 80 Hz Discussion The significant improvement in classification accuracy for 2 out of 50 sampling rates is meaningless, considering that 5 % of all results are expected to be significant by chance. In general, it cannot be stated that classification of AR features can be improved by reducing sample rate, but when tuning a system for an individual subject this is an option to consider. References 1. A. Schl¨ogl, D. Flotzinger, and G. Pfurtscheller, Adaptive autoregressive modeling used for single-trial EEG classification, Biomedizinische Technik, 42:162–167, 1997. 2. D. J. McFarland and J. R. Wolpaw, Sensorimotor rhythm-based brain-computer interface (BCI): model order selection for autoregressive spectral analysis, Journal of Neural Engineering, 5:155–162, 2008. 3. R. Leeb, C. Brunner, G. R. M¨uller-Putz, A. Schl¨ogl and G. Pfurtscheller, BCI Competition IV - Graz data set 2b, http://www.bbci.de/competition/iv/. Acknowledgements This work is supported by the FWF Project “Coupling Measures in BCIs” (P20848N15).

1/--pages