I'm looking for Mathlab code for the perprocessing of multi channels EEG data. What is the best filter I can use to prepare the EEG for the training of deep learning. How to denoising of EEG signals using wavelets or any other tool.? Any code available? Please help.
See similar question "Can i get the matlab code for denoising of Mt data using discrete wavelet transform?" (https://www.researchgate.net/post/can_i_get_the_matlab_code_for_denoising_of_Mt_data_using_discrete_wavelet_transform)
See also "Smoothing and Denoising" (https://www.mathworks.com/help/signal/smoothing-and-denoising.html)
You are asking a very important question, which appears in any signal processing problem, and associated classification/machine learning.
You have a signal s(t), for discrete time t. Your signal may have high frequency components, in other terms, you may see it jump above +M and below -m, where M and m are large positive numbers.
You may want to smooth your signal, you may want to extract features (quasi-stationary) and focus on residual part, which carries changes whatever these changes may be.
Classical methods use ARMA, to extract vicinity properties between say t, t-1, t-2,... t-10 (or maybe up to 100? it depends on the scale of analysis). To be sure call it t, t-1, etc until t-n.
call s^(t) = sum of a(i)*s(t-i) for i=1 to say n. Call e(t)= s(t)-s^(t), this is a prediction residual. The a(i) for i=1 to n have been computed through least square estimation: they are the best fit, minimising the sum of the squares of e(t) for t going from tmin to tmax ([tmin, tmax] is an analysis window, it's the equivalent of training set, with training data s(tmin), s(tmin+1),... s(tmax).
Another method working well for ECG, and any other biological signal with some periodicity (ECG: a man at rest may beat at 60 beats/s, which gives you a period of T=1/60s (f=60Hz). Some very strong athletes beat at 30Hz (T=1/30s), I beat at 120 Hz (T=1/120s) when I run. T varies, but not so much from one cycle to the next.
You want to reduce the signal s(t) from the signal s(t-T). Actually for the kth window of analysis, you fit the best T(k) (estimate quasi-periodicity) and associate a gain g(k) to it, which will be slowly varying (maybe constant on [t0+kT, t0+(k+1)T].
Here the predictor becomes s^(t)= g(k) s(t-Tk), and the residual is e(t)= s(t) -s^(t).
These methods work for ECG, but also for speech (which is produced by the human body, through air going to air pipes, and glottal closure. Look up "speech production"), and for music.
Let me explain on speech: you have resonating signals, highly periodical, e.g. when you say aaaaaa (just a, for a long time).
You have very aperiodic signal, like k, t, which are "plosive" (small explosions produced by the mouth). You also have something in-between like shhhhh, and ssss.
Here you have understood patterns, and the associated models: a vowel has a combination of this long term predictor g(k) s(t-Tk) and the short term predictor sum of a(i) s(t-i).
A plosive has no long-term component (it is highly localised in time) a very sharp e(t) (it is highly dynamic in frequency). You will find details in references on "signal processing for speech".
Now for EEG, you have to associate models (like the a(i), the g(k), Tk) and their value ranges, with patterns and categories you are looking for.
Hope this help.
Try on very small set to check your assumptions...
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Coding