epilepsy master-alpha

at last i find the QMSDP team algorithm for seizure predicting efficient enough.

I'm going to start using it's Q features part in real device temporary and after a success running the algorithm on electrical device, I'll add other QMSDP parts to it.

you can see whole project here: https://github.com/drewabbot/kaggle-seizure-prediction

i can tell a summary of Q features algorithm as following:

using lasso GLM mechanism on following functions:

-Spectrum at six frequency bands: delta (0.1-4Hz), theta (4-8Hz), alpha (8-12Hz), beta (12-30Hz), low-gamma.

(30-70Hz) and high gamma (70-180Hz).

-Spectral edge power of 50% power up to 40Hz.

-Shannon's entropy at dyadic frequency bands.

-Spectrum correlation across channels at dyadic frequency bands.

-Statistical moments: skewness and kurtosis.

this part uses 1 minute windows and 400hz sampling rate.

it will collapses the scores to a single score by mean function.

at the first I'm going to develop a simple code that using GLM method(by matlab lassoglm function) to predict an upcoming seizure. i think it's the most simple kind of machine learning that I'll use following feature that exported from EEG data on it:

1. Spectrum and Shannon's entropy at six frequency bands: delta

(0.1-4Hz), theta (4-8Hz), alpha (8-12Hz), beta (12-30Hz), low-gamma

(30-70Hz) and high gamma (70-180Hz).

2. Spectral edge power of 50% power up to 40Hz.

3. Shannon's entropy at dyadic frequency bands.

4. Spectrum correlation across channels at dyadic frequency bands.

5. Time-series correlation matrix and its eigenvalues.

6. Fractal dimensions.

7. Hjorth parameters: activity, mobility and complexity.

8. Statistical moments: skewness and kurtosis.

I'll use 1 minute window length without overlapping area, but it may be changed by more studies.

this features are used by winner team in kaggle challenge and I'll use them too, because at least we know there is a success experiment of them.

i will use either forest tree and... after completing this scope :)