Hello,
I am trying to implement a frequency-PCA in Matlab/fieldtrip (or Python alternatively).
The idea of this approach is to identify the major frequency components in the EEG data in a data-driven way. Each component is characterized by a dominant frequency and a topography (electrodes that contribute most to it) (like in Barry & De Blasio 2017 for example).
If I understood correctly the steps are:
1. Compute an FFT for each subject and recording
2. Apply PCA on the obtained power spectra across all the subjects and recordings
3. Apply promax or varimax rotation on the obtained coefficients.
I have tried these steps in Matlab, pseudocode below:
powerspectra = [];
for subject in subjects:
for recording in recordings:
cfg = [];
cfg.output = 'pow';
cfg.channel = 'all';
cfg.method = 'mtmfft';
cfg.taper = 'hann';
cfg.keeptrials = 'no';
cfg.foilim = [1 40];
FFT_EEG = ft_freqanalysis(cfg, data_eeg);
powerspectra = [powerspectra; FFT_EEG.powspctrm];
end
end
[coeff,score,latent] = pca(powerspectra);
% Perform Promax rotation on the principal components
rot_matrix = promax(coeff,3);
% apply the rotation to the principal components
rot_coeff = coeff*rot_matrix';
However, the resulting components do not look meaningful (except maybe the first component, see figures attached).
There must be something I'm missing in my implementation? I would greatly appreciate any hint or lead. Thank you
Nicolai Wolpert Based on your pseudocode, it appears like you are on the correct road for developing a frequency-PCA on EEG data. However, some more procedures or tweaks might be taken to improve the outcomes.
One factor to consider is the selection of FFT analysis parameters such as taper and frequency range. Experimenting with various settings may produce better results.
Another factor to consider before using PCA is the normalization of the power spectra. It could be advantageous to normalize the power spectrum by the overall power or the sum of the power spectra's squares.
It's also worth noting that the number of main components to keep for subsequent analysis is a subjective decision that might vary depending on the research question or data collection. Experiment with different numbers of components to find which yields the most relevant results.
Finally, it may be advantageous to depict the topographies of the components in a topographical map in order to have a better understanding of the component's geographical distribution.
Please let me know if you have any more questions!
Qamar Ul Islam Thank you for the recommendations. Unfortunately the results do not make much sense to me, no matter how the listed parameters are chosen. To illustrate: I performed some simulations where I took a typical 1/f power distribution and added a peak in each of 3 virtual "channels" (see figures attached): One peak around 10 Hz in channel Pz, one around 5 Hz in channel Fp1, and one around 30 Hz in channel T3. I simulated 50 recordings where I randomly varied the peak amplitude and spread from one recording to the next. I was expecting that PCA should give me one component with a peak around 10 Hz, one with a peak around 5 Hz and one with a peak around 30 Hz. However, when inspecting the raw components, the first always has all three peaks and the other ones have strange peaks and even negative values (how come?). The rotated components sometimes at least give one clear dominant frequency per component (depends on the iteration), but the peaks are very sharp and also sometimes negative. The number of components and whether normalization was applied or not does not change these results. So since the results already look weird for these clean simulated data I wonder whether there is something one in my procedure.
Concerning the topographical maps, how would you extract these? Run a correlation of the original power spectra in each channel with each of the component power spectra? That would then give a map for each component showing which channel is most similar or contributes most to the components.
Aparna Sathya Murthy Thank you! I had not found an appropriate rotation function in Python so this will be useful.
Hi, please check the methods of the following recent research paper applying PCA for early plant diseases detection:
Article Abaxial leaf surface-mounted multimodal wearable sensor for ...
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Programming Languages
- Python