EEG Analysis

The goal of this dataset is to detect a certain type of pattern (called “spindle”) inside Electroencephalogram (EEG) waves.

Data come from Garches Hospital Sleep Studies Lab (France, Paris area)

They are extracted from a single EEG, mostly from C3-A1 and C4-A1 electrodes

We have used 2 kinds of inputs to read these waves :

• The raw difference in potential measured by each captor
• The frequencies and power spectral densities issued from GeneCycle package periodogram function applied on the result of Fast Fourier Transformation of differences in potential:

raw Difference in potential -> FFT -> periodogram -> frequencies and power spectral densities
The use of frequencies and densities gives much better results.

We know that the spindles must occur in the middle of a certain type of waves (theta waves), so we take as a training set a mixture of spindles and theta waves, and associate the outcome value 1 (as factor) for spindles and value 0 for theta waves.
The 1st part of the algorithm already identified the “spindle candidates” from windows of 1 second in the EEG raw waves and sent these candidates to the “authentication” program, which is the one described. So I used R neural networks (nnet) and Kernel Support Vector Machine (KSVM) for the authentication.

Here is the program with neural net.

In version 2, I added raw EEg data of 30 mn of sleep (extrait_wSleepPage01.csv) and expected patterns (spindles.csv) manually extracted.
If one of you guys has a method to more directly identify spindles from the sleep page, you are welcome !