# First, we will start by generating a noisy simulated power spectrum # ################################################################################################### # Set the frequency range to generate the power spectrum f_range = [1, 50] # Set aperiodic component parameters, as [offset, exponent] ap_params = [20, 2] # Gaussian peak parameters gauss_params = [[10, 1.0, 2.5], [20, 0.8, 2], [32, 0.6, 1]] # Set the level of noise to generate the power spectrum with nlv = 0.1 # Set random seed, for consistency generating simulated data set_random_seed(21) # Create a simulated power spectrum freqs, spectrum = gen_power_spectrum(f_range, ap_params, gauss_params, nlv) ################################################################################################### # Fit an (unconstrained) model, liable to overfit fm = FOOOF() fm.report(freqs, spectrum) ################################################################################################### # # Notice that in the above fit, we are very likely to think that the model has # been overzealous in fitting peaks, and is therefore overfitting. #
def FOOOF(eeg_epoch_full_df):
# Import required code for visualizing example models
from fooof import FOOOF
from fooof.sim.gen import gen_power_spectrum
from fooof.sim.utils import set_random_seed
from fooof.plts.annotate import plot_annotated_model
# Set random seed, for consistency generating simulated data
set_random_seed(10)
# Simulate example power spectra
freqs1, powers1 = gen_power_spectrum([3, 40], [1, 1],
[[10, 0.2, 1.25], [30, 0.15, 2]])
# Initialize power spectrum model objects and fit the power spectra
fm1 = FOOOF(min_peak_height=0.05, verbose=False)
fm1.fit(freqs1, powers1)
plot_annotated_model(fm1, annotate_aperiodic=True)
# Constants
freq_range = [1, 40]
alpha_range = (7, 12)
# Initialize a FOOOF object
fm = FOOOF(peak_width_limits=[1, 8], max_n_peaks=6, min_peak_height=0.4)
# Get the PSD of our EEG Signal
sig = eeg_epoch_full_df['Cz'][0]
freq, psd = getMeanFreqPSD(sig)
fm.add_data(freq, psd, freq_range)
fm.fit()
fm.report()
# Get PSD averages for each channel for each event type (0=left or 1=right)
psd_averages_by_type = {}
for event_type in event_types.keys():
psds_only_one_type = {}
freqs_only_one_type = {}
for i, row in eeg_epoch_full_df[eeg_epoch_full_df["event_type"] ==
event_type].iterrows():
for ch in all_chans:
if ch not in psds_only_one_type:
psds_only_one_type[ch] = list()
freqs_only_one_type[ch] = list()
f, p = getMeanFreqPSD(row[ch])
psds_only_one_type[ch].append(p)
freqs_only_one_type[ch].append(f)
avg_psds_one_type = {}
for ch in all_chans:
psds_only_one_type[ch] = np.array(psds_only_one_type[ch])
avg_psds_one_type[ch] = np.mean(psds_only_one_type[ch], axis=0)
psd_averages_by_type[event_type] = dict(avg_psds_one_type)
# Visualize the parameters in these two classes of C4 activity
fm_left_C4 = FOOOF(peak_width_limits=[1, 8],
max_n_peaks=6,
min_peak_height=0.4)
fm_left_C4.add_data(freqs_only_one_type[eeg_chans[0]][0],
psd_averages_by_type[0]['C4'], freq_range)
fm_left_C4.fit()
fm_left_C4.report()
fm_right_C4 = FOOOF(peak_width_limits=[1, 8],
max_n_peaks=6,
min_peak_height=0.4)
fm_right_C4.add_data(freqs_only_one_type[eeg_chans[0]][0],
psd_averages_by_type[1]['C4'], freq_range)
fm_right_C4.fit()
fm_right_C4.report()
# Calculate central freq, alpha power, and bandwidth for each channel and each trial
# This cell takes a few minutes to run (~8 mins on my computer). There are 3680 trials in the training data.
# Initialize a fooof object
fm = FOOOF(peak_width_limits=[1, 8], max_n_peaks=6, min_peak_height=0.4)
# Some will not have alpha peaks, use these variables to keep track
num_with_alpha = 0
num_without_alpha = 0
fooof_parameters = {}
for i in range(len(eeg_epoch_full_df)):
# Print the trial number every 100 to make sure we're making progress
if i % 100 == 0:
print(i)
for ch in all_chans:
# Determine the key
CF_key = ch + "_alpha_central_freq"
PW_key = ch + "_alpha_power"
BW_key = ch + "_alpha_band_width"
if CF_key not in fooof_parameters:
fooof_parameters[CF_key] = []
if PW_key not in fooof_parameters:
fooof_parameters[PW_key] = []
if BW_key not in fooof_parameters:
fooof_parameters[BW_key] = []
# Calculate the PSD for the desired signal
sig = eeg_epoch_full_df[ch][i]
freq, psd = getMeanFreqPSD(sig)
# Set the frequency and spectral data into the FOOOF model and get peak params
fm.add_data(freq, psd, freq_range)
fm.fit()
peak_params = fm.peak_params_
# Only select the peaks within alpha power
peak_params_alpha = []
for param in peak_params:
if (param[0] > alpha_range[0]) and (param[0] < alpha_range[1]):
peak_params_alpha.append(param)
# Take the average if there are multiple peaks detected, otherwise 0 everything
means = []
if len(peak_params_alpha) > 0:
num_with_alpha += 1
means = np.mean(peak_params_alpha, axis=0)
else:
num_without_alpha += 1
means = [0, 0, 0]
fooof_parameters[CF_key].append(means[0])
fooof_parameters[PW_key].append(means[1])
fooof_parameters[BW_key].append(means[2])
fooof_parameters_df = pd.DataFrame(fooof_parameters)
print("% with alpha:",
num_with_alpha / (num_with_alpha + num_without_alpha))
return fooof_parameters_df