Hello,

I am young researcher working on BCI. I am using common spatial patterns for feature extraction .

I am getting a very low accuracy.

1) I got two classes , each having 60 trials, I separated 42 trails as training and remaining 18 as testing from each class. I have 14 channel headset and sampling rate of 100Hz.

2 ) I obtained the weight matrix (W) using training data only and verified W'*(cov1+cov2)*W=I. So I am sure till this point algorithm is fine.

3 ) I have taken first 3 and last 3 columns of W( 14*6)

4) I projected each trail onto W'(6*14), W'*E where W'=6*14 and E=14*500

5) which gave me s=6*500, then used v= log(var(s))=6*1. v' gave me a 6-d feature vector and labelled it as 1(class1=1 class2=). Same way projected the training data of two classes giving 84 training vectors

6) projected testing data (36 trials) on the same W(which is obtained using training data) and got 36 testing vectors.

7) Used SVM in R software for classification and classification is as low as 45-50% , used other datsets as well but still the accuracy is very low. Can anyone please point me out with the error I made. I am happy to send my code and data if you need.

Any help is very much appreciated.

Kind Regards.

## Spatial Patters and low accuracy

### Re: Spatial Patters and low accuracy

Suhaasbci,

I can only give the general advise to follow these steps:

(1) Visualize the neural correlates, i.e., topographies for alpha, mu, beta and low gamma frequency bands.

(2) Calculate your common spatial patterns feature extraction.

(3) Visualize your common spatial patters filter, i.e., topographies with the weights of your common spatial pattern filter.

(4) Compare the topographies from (1) and (3). There should be some resemblance.

(5) Perform a linear classification on univariate features (i.e., the features of one electrode).

(6) Perform a linear classification on your CSP projected features.

(7) Compare (5) and (6). The results from (6) should be better than any of the results from (5).

(8) Apply the SVM approach on the CSP projected features. The results should be better than they of (6).

Performing this step by step, first verifies the physiological effect, and that this effect can be detected using simple linear regression methods. After you know this, you can compare it to your CSP and SVM approach.

Regards, Peter

I can only give the general advise to follow these steps:

(1) Visualize the neural correlates, i.e., topographies for alpha, mu, beta and low gamma frequency bands.

(2) Calculate your common spatial patterns feature extraction.

(3) Visualize your common spatial patters filter, i.e., topographies with the weights of your common spatial pattern filter.

(4) Compare the topographies from (1) and (3). There should be some resemblance.

(5) Perform a linear classification on univariate features (i.e., the features of one electrode).

(6) Perform a linear classification on your CSP projected features.

(7) Compare (5) and (6). The results from (6) should be better than any of the results from (5).

(8) Apply the SVM approach on the CSP projected features. The results should be better than they of (6).

Performing this step by step, first verifies the physiological effect, and that this effect can be detected using simple linear regression methods. After you know this, you can compare it to your CSP and SVM approach.

Regards, Peter

### Re: Spatial Patters and low accuracy

Sorry for the delay,

Thanks peter got it now.

Thanks peter got it now.

### Who is online

Users browsing this forum: No registered users and 1 guest