[Biosig-general] Good progress

SourceForge Headquarters 225 Broadway Suite 1600 San Diego, CA 92101 +1 (858) 454-5900

Hi schlogl,

I downloaded the datasets 2a and 2b from bci competition iV
So far, i extracted the files and under Matlab I tried to ran demo2.m over
the gdf samples.
I ran on two diferents computers and i got diferent errors on both.
Basically the errors were on the exhaustives matrices operations, in the
step 3 - feature extraction
The computers just ran out of memory such as in the errors bellow.
I reallly thought their processors would burn! :-)

#COMPUTER A: AMD Atlhon XP 1.5 Ghz- 1.5GB  RAM memory

Step 3: feature extraction.
 a: TDP (log(Hjorth)).
??? Error using ==> filter
Out of memory. Type HELP MEMORY for your options.

Error in ==> tdp at 117
         m = filter(B,A,d.^2)./filter(B,A,double(~isnan(d0)));

Error in ==> demo2 at 133
f1 = tdp(s(:,eegchan),p,MODE.UC);

#COMPUTER B: AMD Atlhon 2.0 ghz 64 bits  1GB RAM memory

??? Error using ==> horzcat
Out of memory. Type HELP MEMORY for your options.

Error in ==> tdp at 118
         F = [F, log(m)];

Error in ==> demo2 at 133
f1 = tdp(s(:,eegchan),p,MODE.UC);

#######################################

I debugged every execution of each file using matlab debug tools.

But Suddenly, Eureka!!!

I successfully got the results bellow using the file B0903T.gdf

#################################################################

Step 1: Prepare data.
 a: Load data file B0903T.gdf.
arg[0]=B0903T.gdf
arg[1] IsNumeric
arg[2]=OVERFLOWDETECTION:ON
arg[3]=UCAL:OFF
#info @0117DC98
#0: @0E9B1818 EEG:C3
#1: @0E9B1958 EEG:Cz
#2: @0E9B1A98 EEG:C4
#3: @0E9B1BD8 EOG:ch01
#4: @0E9B1D18 EOG:ch02
#5: @0E9B1E58 EOG:ch03
 b: Extract trigger and classlabels.
Step 2: Preprocessing.
 a: Quality control with histogram analysis [Schloegl et al. 1999].
arg[0]=B0903T.gdf
arg[1] IsNumeric
arg[2]=OVERFLOWDETECTION:OFF
arg[3]=UCAL:ON
#info @0117DC98
#0: @0E9B1818 EEG:C3
#1: @0E9B1958 EEG:Cz
#2: @0E9B1A98 EEG:C4
#3: @0E9B1BD8 EOG:ch01
#4: @0E9B1D18 EOG:ch02
#5: @0E9B1E58 EOG:ch03

EEG2HIST:>Now You can modify the thresholds with mouse clicks.
EEG2HIST:>When you are finished, PRESS ANY KEY on the keyboard.
EEG2HIST:> (make sure the focus is on the figure window). <FINISHED>
 b: Detection of muscle artifacts.
 c: resampling.
 d: reduce EOG artifacts.
 e: spatial filters.
Step 3: feature extraction.
 a: TDP (log(Hjorth)).
 b: Adaptive Autoregressive parameters (AAR).
 c: bandpower.
Step 4: classification.
 b: weights of linear classifier.

ans =

  1.0e+003 *

   (1,1)      -0.0214
   (2,1)       0.0006
   (3,1)       0.0008
   (4,1)      -0.0009
   (5,1)       0.0144
   (6,1)      -0.0029
   (7,1)      -0.0252
   (8,1)      -0.0512
.
.
.

   (9,13)         NaN
  (10,13)         NaN
  (11,13)         NaN
  (12,13)         NaN
  (13,13)         NaN

 c: choosen time segment.

ans =

    5.2040    5.6000

ans =

    5.6040    6.0000

ans =

    8.4040    8.8000

Step 5: classifier.
  Fig 1: results from TDP+LDA.
Classifier: classifier:statistical:lda
Error:     14.4 %
Accuracy:  85.6 %
specific Accuracy: 85.0   86.2
Kappa:     0.71 ± 0.12
I(Wolpaw):  0.41 bit
I(Nykopp):  0.41 bit
I(Continous):     SUM = 0.38  [ 0.38   0.38    ]
STMI:     0.19  [    0.19   0.19 ]
SNR:   0.68   0.68
correlation (parametric): 0.64   0.64
rank correlation:
AUC:   0.90   0.90

ans =

    65    15
     8    72

  Fig 2: results from AAR+LDA results.
Classifier: classifier:statistical:lda
Error:     15.0 %
Accuracy:  85.0 %
specific Accuracy: 84.8   85.2
Kappa:     0.70 ± 0.12
I(Wolpaw):  0.39 bit
I(Nykopp):  0.39 bit
I(Continous):     SUM = 0.39  [ 0.39   0.39    ]
STMI:     0.20  [    0.20   0.20 ]
SNR:   0.72   0.72
correlation (parametric): 0.65   0.65
rank correlation:
AUC:   0.89   0.89

ans =

    67    13
    11    69

  Fig 3: results from BandPower+LDA results.
Classifier: classifier:statistical:lda
Error:     18.1 %
Accuracy:  81.9 %
specific Accuracy: 81.3   82.4
Kappa:     0.64 ± 0.12
I(Wolpaw):  0.32 bit
I(Nykopp):  0.32 bit
I(Continous):     SUM = 0.31  [ 0.31   0.31    ]
STMI:     0.08  [    0.08   0.08 ]
SNR:   0.54   0.54
correlation (parametric): 0.59   0.59
rank correlation:
AUC:   0.87   0.87

ans =

    63    17
    12    68

  Fig 3+: various evaluation criteria [Schlögl et al. 2007] for comparing
different features   Fig 4: Error rate (CC.TSD.ERR)
  Fig 5: correlation coefficient (CC.TSD.r)
  Fig 6: Mutual information (CC.TSD.I)
  Fig 7: Signal-to-Noise ratio (CC.TSD.SNR)
  Fig 8: Area-under-the-ROC-curve (CC.TSD.AUC)
  Fig 9: Accuracy (CC.TSD.ACC00)
  Fig 10: Cohens kappa coefficient (CC.TSD.KAP00)
  Fig 11: Information transfer [Wolpaw] (CC.TSD.I_wolpaw)
  Fig 12: Information transfer [Nykopp] (CC.TSD.I_Nykopp)
  Fig 13: Steepness of mutual information (CC.TSD.STMI)

###################################################################################

I got all the graphics and I they showed the results based on the methods
applied.
And the bottom line is:
 a. where and how do i identify in the graphics the left and right hand
movements?

 b. Furthemore, How to i classify them to return only a single bit answer
such as in the pseudocode:
  if (left_hand_movement)
    return 1
  if (right_hand_movement)
   return 0

I am excited, I am looking forward to read your answer.
cheers,
iuri

ps. i forwarded this email to biosig mailing list just to keep records