Adam Miller, Ryan Martino, Vince Bjazevic

Above shows an illustration of the multi-channel positional layouts for 52 channels.

Method

Above is an illistration of the experiment itself. The way the data was collected from the 8 subjects; each of the subjects were tasked with performing trials followed by consecutive rest periods for each trial. 2 seconds before the task started a green bar appeared on screen indicating the subjects get ready and the equipment would start recording. The subjects underwent 3-4 trials, each trial containing simple sequential subtraction problems. An example being 97 - 4, then 93 - 4 etc. After the experiment, the data complied to produce graphs depicting all 52 channels produced by the EEG. We then took that data and produced code to extract the listed 7 features below. By extracting 7 features from the 52 total channels, our end amount of features extracted was a 1092 total features.

Initially we did not use MatLab's built in function "fitcsvm" which did hamper our progress on the classification model. Below is the "fitcsvm statement.

%fitcsvm statement
mdl = fitcsvm(trainfeatures, trainclasses, 'Standardize', true, 'KernelFunction', 'rbf', 'OptimizeHyperparameters', auto);

During the testing and training phase, seven subjects were used to train the svm and the one who was not included in the training was used to test the model. Initially the subjects being tested were chosen at random, but eventually cycled the testing position consecutively, through use of for loops. This change can attribute to an excessively long runtime. Below is an example plane of optimized hyperparameters for subject two,there are two parameters being tested for, KernelScale and BoxConstraint. It tests combinations (represented by blue dots) and finds the lowest estimated objective function value.

Features

Here is a list of features our code extracted, followed by a method of feature extraction:

Statistical:

1. mean
2. standard deviation
3. skewness
4. kurtosis

Hjorth descriptors:

1. activity
2. mobility
3. complexity

%Feature Extraction Code
datachunk = readings(timestamps(b):timestamps(b+1),:);

            featureActivity = var(datachunk);
            dc1 = datachunk;
            dc2 = datachunk;
            for c = 2:size(datachunk,1)
                dc1(c,:) = datachunk(c,:) - datachunk(c-1,:);
            end
            featureMobility = std(dc1)/std(datachunk);
            for c = 3:size(datachunk,1)
                dc2(c,:) = datachunk(c,:) - 2*datachunk(c-1,:) + datachunk(c-2,:);
            end
            featureFF = (std(dc2)/std(dc1))/(std(dc1)/std(datachunk));
            featureMean = mean(datachunk);
            featureKurtosis = kurtosis(datachunk);
            featureSkewness = skewness(datachunk);
            featureStd = std(datachunk);

Results

Here is the graph showing all 52 channels produced by Subject 2 during trial 1, without the sliding feature implimented.

Subject 2 trial 1, with the sliding window implimented

A table containing all subject and their (sliding/non-sliding) trials

Subject #	Total Trials	Trial 1		Trial 2		Trial 3		Trial 4
Subject 1	3	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding	n/a	n/a
Subject 2	3	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding	n/a	n/a
Subject 3	3	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding	n/a	n/a
Subject 4	4	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding
Subject 5	4	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding
Subject 6	4	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding
Subject 7	4	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding
Subject 8	4	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding	Non-Sliding	Sliding
Subject #	Total Trials	Trial 1		Trial 2		Trial 3		Trail 4

Overall our result, were recorded as follows:

For no sliding window, the Average Accuracy is 0.8550, the Average Fscore is 0.9399, the Accuracy STD is 0.0917, and Fscore STD is 0.0526.

For sliding window, the Average Accuracy is 0.5090, the Average Fscore is 0.7607, the Accuracy STD is 0.0596, the Fscore STD is 0.0662.

Conclusions