关于

DEAP数据集是一个常用的情绪分类公共数据，在日常研究中经常被使用到。如何合理地预处理DEAP数据集，对于后端任务的成功与否，非常重要。本文主要介绍DEAP数据集的预处理流程。

工具

 图片来源：DEAP: A Dataset for Emotion Analysis using Physiological and Audiovisual Signals

 DEAP数据集详细描述：https://www.eecs.qmul.ac.uk/mmv/datasets/deap/doc/tac_special_issue_2011.pdf

方法实现

加载.bdf文件+去除非脑电通道

	bdf_file_name = 's{:02d}.bdf'.format(subject_id)
	bdf_file_path = os.path.join(root_folder, bdf_file_name)
	
	print('Loading .bdf file {}'.format(bdf_file_path))
	raw = mne.io.read_raw_bdf(bdf_file_path, preload=True, verbose=False).load_data()
	ch_names = raw.ch_names
	eeg_channels = ch_names[:N_EEG_electrodes]
	non_eeg_channels = ch_names[N_EEG_electrodes:]
	stim_ch_name = ch_names[-1]
	stim_channels = [ stim_ch_name ]

设置脑电电极规范

raw_copy = raw.copy()
	raw_stim = raw_copy.pick_channels(stim_channels)
	raw.pick_channels(eeg_channels)
	print('Setting montage with BioSemi32 electrode locations')
	biosemi_montage = mne.channels.make_standard_montage(kind='biosemi32', head_size=0.095)
	raw.set_montage(biosemi_montage)

带通滤波器（4-45Hz），陷波滤波器（50Hz）

	print('Applying notch filter (50Hz) and bandpass filter (4-45Hz)')
	raw.notch_filter(np.arange(50, 251, 50), n_jobs=1, fir_design='firwin')
	raw.filter(4, 45, fir_design='firwin')

脑电通道相对于均值重标定

	# Average reference. This is normally added by default, but can also be added explicitly.
	print('Re-referencing all electrodes to the common average reference')
	raw.set_eeg_reference()

获取事件标签

print('Getting events from the status channel')
	events = mne.find_events(raw_stim, stim_channel=stim_ch_name, verbose=True)
	if subject_id<=23:
		# Subject 1-22 and Subjects 23-28 have 48 channels.
		# Subjects 29-32 have 49 channels.
		# For Subjects 1-22 and Subject 23, the stimuli channel has the name 'Status'
		# For Subjects 24-28, the stimuli channel has the name ''
		# For Subjects 29-32, the stimuli channels have the names '-0' and '-1'
		pass
	else:
		# The values of the stimuli channel have to be changed for Subjects 24-32
		# Trigger channel has a non-zero initial value of 1703680 (consider using initial_event=True to detect this event)
		events[:,2] -= 1703680 # subtracting initial value
		events[:,2] = events[:,2] % 65536 # getting modulo with 65536
	
	print('')
	event_IDs = np.unique(events[:,2])
	for event_id in event_IDs:
		col = events[:,2]
		print('Event ID {} : {:05}'.format(event_id, np.sum( 1.0*(col==event_id) ) ) )

获取事件对应信号

inds_new_trial = np.where(events[:,2] == 4)[0]
	events_new_trial = events[inds_new_trial,:]
	baseline = (0, 0)
	print('Epoching the data, into [-5sec, +60sec] epochs')
	epochs = mne.Epochs(raw, events_new_trial, event_id=4, tmin=-5.0, tmax=60.0, picks=eeg_channels, baseline=baseline, preload=True)

ICA去除伪迹

print('Fitting ICA to the epoched data, using {} ICA components'.format(N_ICA))
	ica = ICA(n_components=N_ICA, method='fastica', random_state=23)
	ica.fit(epochs)
	ICA_model_file = os.path.join(ICA_models_folder, 's{:02}_ICA_model.pkl'.format(subject_id))
	with open(ICA_model_file, 'wb') as pkl_file:
		pickle.dump(ica, pkl_file)
	# ica.plot_sources(epochs)
	print('Plotting ICA components')
	fig = ica.plot_components()
	cnt = 1
	for fig_x in fig:
		print(fig_x)
		fig_ICA_path = os.path.join(ICA_components_folder, 's{:02}_ICA_components_{}.png'.format(subject_id, cnt))
		fig_x.savefig(fig_ICA_path)
		cnt += 1
	# Inspect frontal channels to check artifact removal 
	# ica.plot_overlay(raw, picks=['Fp1'])
	# ica.plot_overlay(raw, picks=['Fp2'])
	# ica.plot_overlay(raw, picks=['AF3'])
	# ica.plot_overlay(raw, picks=['AF4'])
	N_excluded_channels = len(ica.exclude)
	print('Excluding {:02} ICA component(s): {}'.format(N_excluded_channels, ica.exclude))
	epochs_clean = ica.apply(epochs.copy())

	#############################

	print('Plotting PSD of epoched data')
	fig = epochs_clean.plot_psd(fmin=4, fmax=45, area_mode='range', average=False, picks=eeg_channels, spatial_colors=True)
	fig_PSD_path = os.path.join(PSD_folder, 's{:02}_PSD.png'.format(subject_id))
	fig.savefig(fig_PSD_path)

	print('Saving ICA epoched data as .pkl file')
	mneraw_pkl_path = os.path.join(mneraw_as_pkl_folder, 's{:02}.pkl'.format(subject_id))
	with open(mneraw_pkl_path, 'wb') as pkl_file:
		pickle.dump(epochs_clean, pkl_file)

	epochs_clean_copy = epochs_clean.copy()

数据降采样和保存

print('Downsampling epoched data to 128Hz')
	epochs_clean_downsampled = epochs_clean_copy.resample(sfreq=128.0)

	print('Plotting PSD of epoched downsampled data')
	fig = epochs_clean_downsampled.plot_psd(fmin=4, fmax=45, area_mode='range', average=False, picks=eeg_channels, spatial_colors=True)
	fig_PSD_path = os.path.join(PSD_folder, 's{:02}_PSD_downsampled.png'.format(subject_id))
	fig.savefig(fig_PSD_path)

	data = epochs_clean.get_data()
	data_downsampled = epochs_clean_downsampled.get_data()
	print('Original epoched data shape: {}'.format(data.shape))
	print('Downsampled epoched data shape: {}'.format(data_downsampled.shape))
	
	###########################################
	
	EEG_channels_table = pd.read_excel(DEAP_EEG_channels_xlsx_path)
	EEG_channels_geneva = EEG_channels_table['Channel_name_Geneva'].values
	channel_pick_indices = []
	print('\nPreparing EEG channel reordering to comply with the Geneva order')
	for (geneva_ch_index, geneva_ch_name) in zip(range(N_EEG_electrodes), EEG_channels_geneva):
		bdf_ch_index = eeg_channels.index(geneva_ch_name)
		channel_pick_indices.append(bdf_ch_index)
		print('Picking source (raw) channel #{:02} to fill target (npy) channel #{:02} | Electrode position: {}'.format(bdf_ch_index + 1, geneva_ch_index + 1, geneva_ch_name))

	ratings = pd.read_csv(ratings_csv_path)
	is_subject =  (ratings['Participant_id'] == subject_id)
	ratings_subj = ratings[is_subject]
	trial_pick_indices = []
	print('\nPreparing EEG trial reordering, from presentation order, to video (Experiment_id) order')
	for i in range(N_trials):
		exp_id = i+1
		is_exp = (ratings['Experiment_id'] == exp_id)
		trial_id = ratings_subj[is_exp]['Trial'].values[0]
		trial_pick_indices.append(trial_id - 1)
		print('Picking source (raw) trial #{:02} to fill target (npy) trial #{:02} | Experiment_id: {:02}'.format(trial_id, exp_id, exp_id))
	
	# Store clean and reordered data to numpy array
	epoch_duration = data_downsampled.shape[-1]
	data_npy = np.zeros((N_trials, N_EEG_electrodes, epoch_duration))
	print('\nStoring the final EEG data in a numpy array of shape {}'.format(data_npy.shape))
	for trial_source, trial_target in zip(trial_pick_indices, range(N_trials)):
		data_trial = data_downsampled[trial_source]
		data_trial_reordered_channels = data_trial[channel_pick_indices,:]
		data_npy[trial_target,:,:] = data_trial_reordered_channels.copy()
	print('Saving the final EEG data in a .npy file')
	np.save(npy_path, data_npy)
	
	print('Raw EEG has been filtered, common average referenced, epoched, artifact-rejected, downsampled, trial-reordered and channel-reordered.')
	print('Finished.')

图片来源： Blog - Ayan's Blog

代码获取

后台私信，请注明文章名称；

相关项目和代码问题，欢迎交流。