METHODS:
The method is based on ARX model fitting of electrode pairs. Each
electrode
signal is fitted by an 8th order autoregressive model (autofit), while
model pairs (there are 64/2 * 63 such pairs) are fitted with 4th order
models - this is termed coupled fit. Prior to this the raw data is band
pass filtered in the alpha, beta and gamma ranges, the final frequency
band
used being 8 to 40Hz.
A first step fits the entire training data, filtered, to auto and
coupled
models for classes 1 and 2. An upper diagonal matrix of coupled fit
strengths is computed, based on finite prediction error (FPE). For a
pair
of electrodes:
FpeRatio= Fpe(CoupledModel i,j, TrainingData)/Sum(Fpe(AutoModel,
TrainingData), electrode i and j)
The matrix of FpeRatio difference between class 1 and 2 is then made
sparse
by applying a threshold of 3 sd(fpeRatioDifference) among classes. The
electrode pairs left (SparseSet) then provide contributions to the
classifier, which simply takes a test trial and simulates models for
class
1 and 2, summing FPE over electrode pairs in the SparseSet. The model
with
the lower FPE sum is the classifier output!
No scaling or special treatment was used for test set.