1. Using Feed Forward NN for EEG Signal Classification Amin Fazel April 2006 Department of Computer Science and Electrical Engineering University of Missouri – Kansas City

2. Thursday, April 27, 2006 CS and EE Department UMKC 2/21 Introduction Features Classifier Results Outline

3. Thursday, April 27, 2006 CS and EE Department UMKC 3/21 Introduction Developed systems –Allow trained subjects to communicate via BCIs –BCI control signal Amplitude of mu (8-12 Hz) or high-beta (18-26) rhythms P300 event-related potential Slow Cortical Potential (SCPs) –Frequency domain information mu and/or beta rhythm amplitude –Time domain information P300 dc potential in the form of SCPs

4. Thursday, April 27, 2006 CS and EE Department UMKC 4/21 Data Set This set pertaining to SCPs in a healthy human subject –Subjects were asked to move a cursor up or down on a computer screen while their SCPs were recorded –6 channels with sampling rate 256Hz –Training Set: 268 trials 168 originated from day I 100 trails from day II –Test Set: 293 trials

5. Thursday, April 27, 2006 CS and EE Department UMKC 5/21 Data Set

6. Thursday, April 27, 2006 CS and EE Department UMKC 6/21 Introduction Features Classifier Results Outline

7. Thursday, April 27, 2006 CS and EE Department UMKC 7/21 Feature Used SCPs features: –Using mean subtraction of the time domain data for each of 6 channels To prevent dc offset effect

8. Thursday, April 27, 2006 CS and EE Department UMKC 8/21 Mean Subtraction

9. Thursday, April 27, 2006 CS and EE Department UMKC 9/21 Feature Extraction Linear Predictive Coding (LPC) estimating the parameters of the current signal sample by using the parameter values from linear combinations of past signal samples

10. Thursday, April 27, 2006 CS and EE Department UMKC 10/21 LPC FE - Framming Frame Blocking LPC Analysis s (n) Feature vector NM N Overlap of 50% Framing window = 0.5 sec 13 frames 896 sample or 3.5 sec per channel

11. Thursday, April 27, 2006 CS and EE Department UMKC 11/21 Features: LPC Linear Predictive Coding (LPC) provides – low-dimension representation of signal at one frame – “analytically tractable” method LPC models signal as approximate linear combination of previous p samples: where a 1, a 2, … a p are constant for each frame of signal.

12. Thursday, April 27, 2006 CS and EE Department UMKC 12/21 Linear Predictive Model LPC coefficients are a filter H(z) s(n) u(n) G

13. Thursday, April 27, 2006 CS and EE Department UMKC 13/21 LPC extraction Given a window of signal samples, the first p+1 terms of the autocorrelation sequence are calculated from where i = 0,..,p. The filter coefficients are then computed recursively using a set of auxiliary coefficients k i which can be interpreted as the reflection coefficients and the prediction error E which is initially equal to r 0. Let and be the reflection and filter coefficients for a filter of order i-1, then a filter of order i can be calculated in three steps.

14. Thursday, April 27, 2006 CS and EE Department UMKC 14/21 LPC extraction Firstly, a new set of reflection coefficients are calculated for j = 1,.., i-1 and Secondly, E is updated Finally, new filter coefficients are computed for j = 1,…,i-1 and This process is repeated from i=1 through to the required filter order i=p.

15. Thursday, April 27, 2006 CS and EE Department UMKC 15/21 Introduction Feature Extraction Classifier Results Outline

16. Thursday, April 27, 2006 CS and EE Department UMKC 16/21 Feed Forward NN Multilayer feed forward networks are universal approximators (Hornik 1989) –as few as one hidden layer However, in practice, things are not so simple

17. Thursday, April 27, 2006 CS and EE Department UMKC 17/21 Feed Forward NN As we have –13 frames –12 LPC features for each –6 channels So –936 PEs in the input layer –10, 40 PEs in the hidden layer –1 PE in the output layer

18. Thursday, April 27, 2006 CS and EE Department UMKC 18/21 Introduction Features Classifier Results Outline

19. Thursday, April 27, 2006 CS and EE Department UMKC 19/21 Results Number of Hidden Layer Using validation data set Test Set (% Mean) Test Set (% Best) 10No80.2282.46 10Yes63.4372.39 40Yes62.6266.42 100Yes65.4266.42 500Yes64.9967.91 100No79.1780.22 Mean data subtraction was applied. Features: 12 order LPC

20. Thursday, April 27, 2006 CS and EE Department UMKC 20/21 Introduction Feature s Classifier Results Outline

21. Thanks for your patience !