def analyzeOffline(subjID):
'''
'''
# Initialize conditions for preprocessing of epochs (preproc1epoch from EEG_analysis_RT)
reject_ch = 1 # Rejection of nine predefined channels
reject = None # Rejection of channels, either manually defined or based on MNE analysis
mne_reject = 0
flat = None # Input for MNE rejection
bad_channels = None # Input for manual rejection of channels
opt_detrend = 1 # Temporal EEG detrending (linear)
if reject_ch == 1:
n_channels = 23
if reject_ch == 0:
n_channels = 32
# Initialize conditions for SSP rejection (applySSP from EEG_analysis_RT)
threshold = 0.1 # SSP projection variance threshold
# Initialize dictionary for saving outputs
d = {}
d['subjID'] = subjID
EEGfile, markerFile, idxFile, alphaFile, n_it = findSubjectFiles(
subjID, manual_n_it=None)
#%% Extract epochs from EEG data
prefilter = 0
if subjID in ['07', '08', '11']:
n_samples_fs500 = 550 # Number of samples to extract for each epoch, sampling frequency 500
n_samples_fs100 = int(
550 / 5) # Number of samples, sampling frequency 100 (resampled)
else:
n_samples_fs500 = 450
n_samples_fs100 = int(450 / 5)
e = extractEpochs_tmin(EEGfile,
markerFile,
prefilter=prefilter,
marker1=0,
n_samples=n_samples_fs500)
cat = extractCat(idxFile, exp_type='fused')
# MNE info files
info_fs500 = create_info_mne(reject_ch=0, sfreq=500)
info_fs100 = create_info_mne(reject_ch=1, sfreq=100)
# Make split. Shuffle the pool of e+cats.
fourteen = np.random.choice(cv, 14, replace=False)
split1 = fourteen[0:7]
split2 = fourteen[7:14]
#%% Run NF RT offline analysis
def runRealTime(e):
'''
Run real-time pipeline for one participant based on inputted e and cat.
Output a coef_lst, and make this into
'''
coef_lst = []
stable_blocks0 = e[:600, :, :] # Fi+st run
stable_blocks1 = np.zeros((600, n_channels, n_samples_fs100))
y = np.array([int(x) for x in cat])
y_run = y[:600]
pred_prob_train = np.zeros((n_it * 600, 2))
pred_prob_test = np.zeros((n_it * 200, 2)) # Prediction probability
pred_prob_test_corr = np.zeros(
(n_it * 200,
2)) # Prediction probability, corrected for classifier bias
alpha_test = np.zeros((n_it * 200))
clf_output_test = np.zeros((n_it * 200))
train_acc = np.zeros(n_it)
c_test = 0
c = 0
offset = 0
y_pred_test1 = np.zeros(5 * 200)
y_test_feedback = np.concatenate(
(y[600:800], y[1000:1200], y[1400:1600], y[1800:2000], y[2200:2400]))
Y_train = np.zeros((n_it, 600))
Y_run = np.zeros((n_it, 600))
val = 1 # Offset validation
offset_pred = 0 # Offset prediction
offset_Pred = []
# Score = np.zeros((n_it,3))
epochs_fb = np.zeros((200, 23, n_samples_fs100)) # Epochs feedback
for b in range(n_it):
for t in range(stable_blocks0.shape[0]):
epoch = stable_blocks0[t, :, :]
epoch = preproc1epoch(epoch,
info_fs500,
SSP=False,
reject=reject,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=flat,
bad_channels=bad_channels,
opt_detrend=opt_detrend)
stable_blocks1[t + offset, :, :] = epoch
c += 1
projs1, stable_blocksSSP1 = applySSP(
stable_blocks1, info_fs100,
threshold=threshold) # Apply SSP on stable blocks
# Average training blocks after SSP
stable_blocksSSP1 = average_stable(stable_blocksSSP1)
if val == 1: # Test for offset classifier bias
if b > 0: # Runs after first run
stable_blocksSSP1_append = np.append(stable_blocksSSP1,
epochs_fb,
axis=0)
fb_start = (b - 1) * 200
y_run_append = np.append(y_run,
y_test_feedback[fb_start:fb_start +
200],
axis=0)
clf, offset_pred = trainLogReg_cross2(stable_blocksSSP1_append,
y_run_append)
else:
clf, offset_pred = trainLogReg_cross(stable_blocksSSP1,
y_run) # First run
offset_Pred.append(offset_pred)
#Score[b,:]=score
print('Offset estimated to ' + str(offset_pred))
else:
clf = trainLogReg(stable_blocksSSP1, y_run)
Y_run[b, :] = y_run
y_pred_train1 = np.zeros(len(y_run))
offset_pred = np.min([np.max([offset_pred, -0.25]), 0.25
]) / 2 # Limit offset correction to abs(0.125)
# Training accuracy based on the RT classifier
for t in range(len(y_run)):
epoch_t = stable_blocksSSP1[t, :, :]
pred_prob_train[t, :], y_pred_train1[t] = testEpoch(clf, epoch_t)
Y_train[b, :] = y_pred_train1
train_acc[b] = len(
np.where(np.array(y_pred_train1[0:len(y_run)]) == np.array(y_run))
[0]) / len(y_run)
stable_blocks1 = stable_blocks1[200:, :, :]
stable_blocks1 = np.concatenate(
(stable_blocks1, np.zeros((200, n_channels, n_samples_fs100))),
axis=0)
s_begin = 800 + b * 400
offset = 400 # Indexing offset for EEG data and y
stable_blocks0 = e[s_begin:s_begin + 200, :, :]
y_run = np.concatenate((y_run[200:], y[s_begin:s_begin + 200]))
# Save clf _coefs
coefs = clf.coef_
coef_r = np.reshape(coefs, [23, n_samples_fs100])
coef_lst.append(coef_r)
# Test accuracy of RT epochs
for t in range(200):
print('Testing epoch number: ', c)
epoch = e[c, :, :]
epoch = preproc1epoch(epoch,
info_fs500,
projs=projs1,
SSP=True,
reject=reject,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=flat,
bad_channels=bad_channels,
opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch + epoch_prev) / 2
pred_prob_test[c_test, :], y_pred_test1[c_test] = testEpoch(
clf, epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test, 0] = np.min(
[np.max([pred_prob_test[c_test, 0] + offset_pred, 0]), 1])
pred_prob_test_corr[c_test, 1] = np.min(
[np.max([pred_prob_test[c_test, 1] - offset_pred, 0]), 1])
clf_output = pred_prob_test_corr[
c_test, int(y[c])] - pred_prob_test_corr[c_test,
int(y[c] - 1)]
clf_output_test[
c_test] = clf_output # Save classifier output for correlation checks
alpha_test[c_test] = sigmoid(
clf_output
) # Convert corrected classifier output to an alpha value using a sigmoid transfer function
epoch_prev = epoch
epochs_fb[t, :, :] = epoch
c += 1
c_test += 1
above_chance_offline = len(
np.where((np.array(alpha_test[:c_test]) > 0.5))[0]) / len(
alpha_test[:c_test])
print('Above chance alpha (corrected): ' + str(above_chance_offline))
score = metrics.accuracy_score(y_test_feedback[:c_test],
y_pred_test1[:c_test])
print('Accuracy score (uncorrected): ' + str(score))
# Test score per run
a_per_run = np.zeros((n_it))
score_per_run = np.zeros((n_it))
j = 0
for run in range(n_it):
score_per_run[run] = metrics.accuracy_score(
y_test_feedback[run * 200:(run + 1) * 200],
y_pred_test1[run * 200:(run + 1) * 200])
a_per_run[run] = (
len(np.where((np.array(alpha_test[j:j + 200]) > 0.5))[0]) / 200)
j += 200
print('Alpha, corrected, above chance per run: ' + str(a_per_run))
print('Score, uncorrected, per run: ' + str(score_per_run))
d['RT_test_acc_corr'] = above_chance_offline
d['RT_test_acc_corr_run'] = a_per_run
d['RT_coefs'] = coef_lst
#%%
# Test score per block
a_per_block = np.zeros((n_it * 4))
score_per_block = np.zeros((n_it * 4))
j = 0
for run in range(n_it * 4):
score_per_block[run] = metrics.accuracy_score(
y_test_feedback[run * 50:(run + 1) * 50],
y_pred_test1[run * 50:(run + 1) * 50])
a_per_block[run] = (
len(np.where((np.array(alpha_test[j:j + 50]) > 0.5))[0]) / 50)
j += 50
d['alpha_per_block'] = a_per_block
#%% Save pckl file
pkl_arr = [d]
print('Finished running test and train analyses for subject: ' +
str(subjID))
# PICKLE TIME
fname = '18Jun_subj_' + str(subjID) + '.pkl'
with open(fname, 'wb') as fout:
pickle.dump(pkl_arr, fout)
for t in range(stable_blocks0.shape[0]):
epoch = stable_blocks0[t, :, :]
epoch = preproc1epoch(
epoch,
info_fs500,
SSP=False,
reject=None,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=None,
bad_channels=bad_channels,
opt_detrend=opt_detrend) #stable_blocks.shape[0])
stable_blocks1[t + offset, :, :] = epoch
c += 1
projs1, stable_blocksSSP1 = applySSP(
stable_blocks1, info_fs100,
threshold=threshold) # Apply SSP on stable blocks
# Average training blocks after SSP
stable_blocksSSP1 = average_stable(stable_blocksSSP1)
if val == 1: # Test for offset classifier bias
if b > 0: # Runs after first run
stable_blocksSSP1_append = np.append(stable_blocksSSP1,
epochs_fb,
axis=0)
fb_start = (b - 1) * 200
y_run_append = np.append(y_run,
y_test_feedback[fb_start:fb_start + 200],
axis=0)
clf, offset_pred = trainLogReg_cross2(stable_blocksSSP1_append,
def analyzeOffline(subjID):
'''
'''
# Initialize conditions for preprocessing of epochs (preproc1epoch from EEG_analysis_RT)
reject_ch = 1 # Rejection of nine predefined channels
reject = None # Rejection of channels, either manually defined or based on MNE analysis
mne_reject = 0
flat = None # Input for MNE rejection
bad_channels = None # Input for manual rejection of channels
opt_detrend = 1 # Temporal EEG detrending (linear)
if reject_ch == 1:
n_channels = 23
if reject_ch == 0:
n_channels = 32
# Initialize conditions for SSP rejection (applySSP from EEG_analysis_RT)
threshold = 0.1 # SSP projection variance threshold
# Initialize dictionary for saving outputs
d = {}
d['subjID'] = subjID
plot_MNE = 0
EEGfile, markerFile, idxFile, alphaFile, n_it = findSubjectFiles(
subjID, manual_n_it=None)
#%% Test RT alpha per run
alpha, marker = extractAlpha(alphaFile)
above_chance = len(np.where((np.array(alpha) > 0.5))[0]) / len(alpha)
alpha_per_run = np.zeros((n_it))
j = 0
for ii in range(n_it):
alpha_per_run[ii] = len(
np.where((np.array(alpha[j:j + 200]) > 0.5))[0]) / 200
j += 200
d['ALPHA_fromfile_overall'] = above_chance
d['ALPHA_fromfile_run'] = alpha_per_run
#%% Extract epochs from EEG data
prefilter = 0
if subjID in ['07', '08', '11']:
n_samples_fs500 = 550 # Number of samples to extract for each epoch, sampling frequency 500
n_samples_fs100 = int(
550 / 5) # Number of samples, sampling frequency 100 (resampled)
else:
n_samples_fs500 = 450
n_samples_fs100 = int(450 / 5)
e = extractEpochs_tmin(EEGfile,
markerFile,
prefilter=prefilter,
marker1=0,
n_samples=n_samples_fs500)
cat = extractCat(idxFile, exp_type='fused')
# MNE info files
info_fs500 = create_info_mne(reject_ch=0, sfreq=500)
info_fs100 = create_info_mne(reject_ch=1, sfreq=100)
#%% Run NF RT offline analysis
coef_lst = []
stable_blocks0 = e[:600, :, :] # Fi+st run
stable_blocks1 = np.zeros((600, n_channels, n_samples_fs100))
y = np.array([int(x) for x in cat])
y_run = y[:600]
pred_prob_train = np.zeros((n_it * 600, 2))
pred_prob_test = np.zeros((n_it * 200, 2)) # Prediction probability
pred_prob_test_corr = np.zeros(
(n_it * 200,
2)) # Prediction probability, corrected for classifier bias
alpha_test = np.zeros((n_it * 200))
clf_output_test = np.zeros((n_it * 200))
train_acc = np.zeros(n_it)
c_test = 0
c = 0
offset = 0
y_pred_test1 = np.zeros(5 * 200)
y_test_feedback = np.concatenate(
(y[600:800], y[1000:1200], y[1400:1600], y[1800:2000], y[2200:2400]))
Y_train = np.zeros((n_it, 600))
Y_run = np.zeros((n_it, 600))
val = 1 # Offset validation
offset_pred = 0 # Offset prediction
offset_Pred = []
# Score = np.zeros((n_it,3))
epochs_fb = np.zeros((200, 23, n_samples_fs100)) # Epochs feedback
for b in range(n_it):
for t in range(stable_blocks0.shape[0]):
epoch = stable_blocks0[t, :, :]
epoch = preproc1epoch(epoch,
info_fs500,
SSP=False,
reject=reject,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=flat,
bad_channels=bad_channels,
opt_detrend=opt_detrend)
stable_blocks1[t + offset, :, :] = epoch
c += 1
projs1, stable_blocksSSP1 = applySSP(
stable_blocks1, info_fs100,
threshold=threshold) # Apply SSP on stable blocks
# Average training blocks after SSP
stable_blocksSSP1 = average_stable(stable_blocksSSP1)
if val == 1: # Test for offset classifier bias
if b > 0: # Runs after first run
stable_blocksSSP1_append = np.append(stable_blocksSSP1,
epochs_fb,
axis=0)
fb_start = (b - 1) * 200
y_run_append = np.append(y_run,
y_test_feedback[fb_start:fb_start +
200],
axis=0)
clf, offset_pred = trainLogReg_cross2(stable_blocksSSP1_append,
y_run_append)
else:
clf, offset_pred = trainLogReg_cross(stable_blocksSSP1,
y_run) # First run
offset_Pred.append(offset_pred)
#Score[b,:]=score
print('Offset estimated to ' + str(offset_pred))
else:
clf = trainLogReg(stable_blocksSSP1, y_run)
Y_run[b, :] = y_run
y_pred_train1 = np.zeros(len(y_run))
offset_pred = np.min([np.max([offset_pred, -0.25]), 0.25
]) / 2 # Limit offset correction to abs(0.125)
# Training accuracy based on the RT classifier
for t in range(len(y_run)):
epoch_t = stable_blocksSSP1[t, :, :]
pred_prob_train[t, :], y_pred_train1[t] = testEpoch(clf, epoch_t)
Y_train[b, :] = y_pred_train1
train_acc[b] = len(
np.where(np.array(y_pred_train1[0:len(y_run)]) == np.array(y_run))
[0]) / len(y_run)
stable_blocks1 = stable_blocks1[200:, :, :]
stable_blocks1 = np.concatenate(
(stable_blocks1, np.zeros((200, n_channels, n_samples_fs100))),
axis=0)
s_begin = 800 + b * 400
offset = 400 # Indexing offset for EEG data and y
stable_blocks0 = e[s_begin:s_begin + 200, :, :]
y_run = np.concatenate((y_run[200:], y[s_begin:s_begin + 200]))
# Save clf _coefs
coefs = clf.coef_
coef_r = np.reshape(coefs, [23, n_samples_fs100])
coef_lst.append(coef_r)
# Test accuracy of RT epochs
for t in range(200):
print('Testing epoch number: ', c)
epoch = e[c, :, :]
epoch = preproc1epoch(epoch,
info_fs500,
projs=projs1,
SSP=True,
reject=reject,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=flat,
bad_channels=bad_channels,
opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch + epoch_prev) / 2
pred_prob_test[c_test, :], y_pred_test1[c_test] = testEpoch(
clf, epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test, 0] = np.min(
[np.max([pred_prob_test[c_test, 0] + offset_pred, 0]), 1])
pred_prob_test_corr[c_test, 1] = np.min(
[np.max([pred_prob_test[c_test, 1] - offset_pred, 0]), 1])
clf_output = pred_prob_test_corr[
c_test, int(y[c])] - pred_prob_test_corr[c_test,
int(y[c] - 1)]
clf_output_test[
c_test] = clf_output # Save classifier output for correlation checks
alpha_test[c_test] = sigmoid(
clf_output
) # Convert corrected classifier output to an alpha value using a sigmoid transfer function
epoch_prev = epoch
epochs_fb[t, :, :] = epoch
c += 1
c_test += 1
above_chance_offline = len(
np.where((np.array(alpha_test[:c_test]) > 0.5))[0]) / len(
alpha_test[:c_test])
print('Above chance alpha (corrected): ' + str(above_chance_offline))
score = metrics.accuracy_score(y_test_feedback[:c_test],
y_pred_test1[:c_test])
print('Accuracy score (uncorrected): ' + str(score))
# Test score per run
a_per_run = np.zeros((n_it))
score_per_run = np.zeros((n_it))
j = 0
for run in range(n_it):
score_per_run[run] = metrics.accuracy_score(
y_test_feedback[run * 200:(run + 1) * 200],
y_pred_test1[run * 200:(run + 1) * 200])
a_per_run[run] = (
len(np.where((np.array(alpha_test[j:j + 200]) > 0.5))[0]) / 200)
j += 200
print('Alpha, corrected, above chance per run: ' + str(a_per_run))
print('Score, uncorrected, per run: ' + str(score_per_run))
d['RT_test_acc_corr'] = above_chance_offline
d['RT_test_acc_corr_run'] = a_per_run
d['RT_coefs'] = coef_lst
#%% Analyze RT session block-wise
d['ALPHA_test'] = alpha_test
d['CLFO_test'] = clf_output_test
#%% Stable blocks - train on first run.
c_test = 0
no_sb_fb = (4 * n_it) - 4 # Number stable blocks in fb run minus train run
block_len = 50
pred_prob_test = np.zeros(
(no_sb_fb * block_len,
2)) # Prediction probability test. Block length of 50 trials
pred_prob_test_corr = np.zeros(
(no_sb_fb * block_len,
2)) # Prediction probability test, corrected for bias
alpha_test = np.zeros(
(no_sb_fb * block_len)) # Alpha values for stable blocks
stable_blocks_fbrun = np.concatenate([
e[400 + n * 400:600 + n * 400] for n in range(n_it)
]) # Stable blocks feedback run
y_stable_blocks_fbrun = np.concatenate(
[y[400 + n * 400:600 + n * 400] for n in range(n_it)])
y_pred = np.zeros(no_sb_fb * block_len)
# Only use the stable blocks from fb runs
for r in range(n_it - 1): # 5 runs - 1
print('Run no: ', r)
# First fb run
stable_blocks_train = stable_blocks_fbrun[:200]
y_train = y_stable_blocks_fbrun[:200]
val_indices = range((r + 1) * 200,
((r + 1) * 200) + 200) # Validation block index
stable_blocks_val = stable_blocks_fbrun[val_indices]
y_val = y_stable_blocks_fbrun[val_indices]
stable_blocks_train_prep = np.zeros(
(len(y_train), 23, n_samples_fs100))
for t in range(stable_blocks_train.shape[0]):
epoch = stable_blocks_train[t, :, :]
epoch = preproc1epoch(epoch,
info_fs500,
SSP=False,
reject=reject,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=flat,
bad_channels=bad_channels,
opt_detrend=opt_detrend)
stable_blocks_train_prep[t, :, :] = epoch
projs1, stable_blocksSSP_train = applySSP(stable_blocks_train_prep,
info_fs100,
threshold=threshold)
# Average after SSP correction
stable_blocksSSP_train = average_stable(stable_blocksSSP_train)
clf, offset_pred = trainLogReg_cross_offline(
stable_blocksSSP_train, y_train) #cur. in EEG_classification
offset_pred = np.min([np.max([offset_pred, -0.25]), 0.25]) / 2
# Test epochs in validation run. Preprocessing and testing epoch-wise
for t in range(len(val_indices)):
epoch = stable_blocks_val[t, :, :]
epoch = preproc1epoch(epoch,
info_fs500,
projs=projs1,
SSP=True,
reject=reject,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=flat,
bad_channels=bad_channels,
opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch + epoch_prev) / 2
pred_prob_test[c_test, :], y_pred[c_test] = testEpoch(clf, epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test, 0] = np.min(
[np.max([pred_prob_test[c_test, 0] + offset_pred, 0]), 1])
pred_prob_test_corr[c_test, 1] = np.min(
[np.max([pred_prob_test[c_test, 1] - offset_pred, 0]), 1])
clf_output = pred_prob_test_corr[
c_test, int(y_val[t])] - pred_prob_test_corr[c_test,
int(y_val[t] - 1)]
alpha_test[c_test] = sigmoid(clf_output)
epoch_prev = epoch
c_test += 1
print('No c_test: ' + str(c_test))
above_chance_train = len(
np.where((np.array(alpha_test[:c_test]) > 0.5))[0]) / len(
alpha_test[:c_test])
print('Above chance alpha train (corrected): ' + str(above_chance_train))
score = metrics.accuracy_score(y_stable_blocks_fbrun[200:], y_pred)
d['LORO_first_acc_corr'] = above_chance_train
d['LORO_first_acc_uncorr'] = score
#%% Stable blocks - train on last run.
c_test = 0
pred_prob_test = np.zeros(
(no_sb_fb * block_len,
2)) # Prediction probability test. Block length of 50 trials
pred_prob_test_corr = np.zeros(
(no_sb_fb * block_len,
2)) # Prediction probability test, corrected for bias
alpha_test = np.zeros(
(no_sb_fb * block_len)) # Alpha values for stable blocks
stable_blocks_fbrun = np.concatenate([
e[400 + n * 400:600 + n * 400] for n in range(n_it)
]) # Stable blocks feedback run
y_stable_blocks_fbrun = np.concatenate(
[y[400 + n * 400:600 + n * 400] for n in range(n_it)])
y_pred = np.zeros(no_sb_fb * block_len)
for r in range(n_it - 1): # 5 runs - 1
print('Run no: ', r)
# Last fb run
stable_blocks_train = stable_blocks_fbrun[800:]
y_train = y_stable_blocks_fbrun[800:]
val_indices = range((r) * 200,
((r) * 200) + 200) # Validation block index
stable_blocks_val = stable_blocks_fbrun[val_indices]
y_val = y_stable_blocks_fbrun[val_indices]
stable_blocks_train_prep = np.zeros(
(len(y_train), 23, n_samples_fs100))
for t in range(stable_blocks_train.shape[0]):
epoch = stable_blocks_train[t, :, :]
epoch = preproc1epoch(epoch,
info_fs500,
SSP=False,
reject=reject,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=flat,
bad_channels=bad_channels,
opt_detrend=opt_detrend)
stable_blocks_train_prep[t, :, :] = epoch
projs1, stable_blocksSSP_train = applySSP(stable_blocks_train_prep,
info_fs100,
threshold=threshold)
# Average after SSP correction
stable_blocksSSP_train = average_stable(stable_blocksSSP_train)
clf, offset_pred = trainLogReg_cross_offline(
stable_blocksSSP_train, y_train) #cur. in EEG_classification
offset_pred = np.min([np.max([offset_pred, -0.25]), 0.25]) / 2
# Test epochs in validation run. Preprocessing and testing epoch-wise
for t in range(len(val_indices)):
epoch = stable_blocks_val[t, :, :]
epoch = preproc1epoch(epoch,
info_fs500,
projs=projs1,
SSP=True,
reject=reject,
mne_reject=mne_reject,
reject_ch=reject_ch,
flat=flat,
bad_channels=bad_channels,
opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch + epoch_prev) / 2
pred_prob_test[c_test, :], y_pred[c_test] = testEpoch(clf, epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test, 0] = np.min(
[np.max([pred_prob_test[c_test, 0] + offset_pred, 0]), 1])
pred_prob_test_corr[c_test, 1] = np.min(
[np.max([pred_prob_test[c_test, 1] - offset_pred, 0]), 1])
clf_output = pred_prob_test_corr[
c_test, int(y_val[t])] - pred_prob_test_corr[c_test,
int(y_val[t] - 1)]
alpha_test[c_test] = sigmoid(clf_output)
epoch_prev = epoch
c_test += 1
print('No c_test: ' + str(c_test))
above_chance_train = len(
np.where((np.array(alpha_test[:c_test]) > 0.5))[0]) / len(
alpha_test[:c_test])
print('Above chance alpha train (corrected): ' + str(above_chance_train))
score = metrics.accuracy_score(y_stable_blocks_fbrun[:800], y_pred)
d['LORO_last_acc_corr'] = above_chance_train
d['LORO_last_acc_uncorr'] = score
#%% Save pckl file
pkl_arr = [d]
print('Finished running test and train analyses for subject: ' +
str(subjID))
# PICKLE TIME
fname = '14Jun_subj_' + str(subjID) + '.pkl'
with open(fname, 'wb') as fout:
pickle.dump(pkl_arr, fout)
def analyzeOffline(subjID):
'''
'''
# Initialize conditions for preprocessing of epochs (preproc1epoch from EEG_analysis_RT)
reject_ch = 1 # Rejection of nine predefined channels
reject = None # Rejection of channels, either manually defined or based on MNE analysis
mne_reject = 0
flat = None # Input for MNE rejection
bad_channels = None # Input for manual rejection of channels
opt_detrend = 1 # Temporal EEG detrending (linear)
if reject_ch == 1:
n_channels = 23
if reject_ch == 0:
n_channels = 32
# Initialize conditions for SSP rejection (applySSP from EEG_analysis_RT)
threshold = 0.1 # SSP projection variance threshold
# Initialize dictionary for saving outputs
d = {}
d['subjID'] = subjID
plot_MNE = 0
EEGfile, markerFile, idxFile, alphaFile, n_it = findSubjectFiles(subjID,manual_n_it=None)
#%% Test RT alpha per run
alpha,marker = extractAlpha(alphaFile)
above_chance = len(np.where((np.array(alpha)>0.5))[0])/len(alpha)
alpha_per_run = np.zeros((n_it))
j = 0
for ii in range(n_it):
alpha_per_run[ii] = len(np.where((np.array(alpha[j:j+200])>0.5))[0])/200
j += 200
d['ALPHA_fromfile_overall'] = above_chance
d['ALPHA_fromfile_run'] = alpha_per_run
#%% Extract epochs from EEG data
prefilter = 0
if subjID in ['07','08','11']:
n_samples_fs500 = 550 # Number of samples to extract for each epoch, sampling frequency 500
n_samples_fs100 = int(550/5) # Number of samples, sampling frequency 100 (resampled)
else:
n_samples_fs500 = 450
n_samples_fs100 = int(450/5)
e = extractEpochs_tmin(EEGfile,markerFile,prefilter=prefilter,marker1=0,n_samples=n_samples_fs500)
cat = extractCat(idxFile,exp_type='fused')
# MNE info files
info_fs500 = create_info_mne(reject_ch=0,sfreq=500)
info_fs100 = create_info_mne(reject_ch=1,sfreq=100)
#%% Run NF RT offline analysis
stable_blocks0 = e[:600,:,:] # Fi+st run
stable_blocks1 = np.zeros((600,n_channels,n_samples_fs100))
y = np.array([int(x) for x in cat])
y_run = y[:600]
pred_prob_train = np.zeros((n_it*600,2))
pred_prob_test = np.zeros((n_it*200,2)) # Prediction probability
pred_prob_test_corr = np.zeros((n_it*200,2)) # Prediction probability, corrected for classifier bias
alpha_test = np.zeros((n_it*200))
clf_output_test = np.zeros((n_it*200))
train_acc = np.zeros(n_it)
c_test = 0
c = 0
offset = 0
y_pred_test1 = np.zeros(5*200)
y_test_feedback = np.concatenate((y[600:800],y[1000:1200],y[1400:1600],y[1800:2000],y[2200:2400]))
Y_train = np.zeros((n_it,600))
Y_run = np.zeros((n_it,600))
val = 1 # Offset validation
offset_pred = 0 # Offset prediction
offset_Pred = []
# Score = np.zeros((n_it,3))
epochs_fb = np.zeros((200,23,n_samples_fs100)) # Epochs feedback
for b in range(n_it):
for t in range(stable_blocks0.shape[0]):
epoch = stable_blocks0[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
stable_blocks1[t+offset,:,:] = epoch
c += 1
projs1,stable_blocksSSP1 = applySSP(stable_blocks1,info_fs100,threshold=threshold) # Apply SSP on stable blocks
# Average training blocks after SSP
stable_blocksSSP1 = average_stable(stable_blocksSSP1)
if val == 1: # Test for offset classifier bias
if b > 0: # Runs after first run
stable_blocksSSP1_append = np.append(stable_blocksSSP1,epochs_fb,axis=0)
fb_start = (b-1)*200
y_run_append = np.append(y_run,y_test_feedback[fb_start:fb_start+200],axis=0)
clf,offset_pred = trainLogReg_cross2(stable_blocksSSP1_append,y_run_append)
else:
clf,offset_pred = trainLogReg_cross(stable_blocksSSP1,y_run) # First run
offset_Pred.append(offset_pred)
#Score[b,:]=score
print('Offset estimated to '+str(offset_pred))
else:
clf = trainLogReg(stable_blocksSSP1,y_run)
Y_run[b,:]=y_run
y_pred_train1 = np.zeros(len(y_run))
offset_pred = np.min([np.max([offset_pred,-0.25]),0.25])/2 # Limit offset correction to abs(0.125)
# Training accuracy based on the RT classifier
for t in range(len(y_run)):
epoch_t = stable_blocksSSP1[t,:,:]
pred_prob_train[t,:],y_pred_train1[t] = testEpoch(clf,epoch_t)
Y_train[b,:]=y_pred_train1
train_acc[b] = len(np.where(np.array(y_pred_train1[0:len(y_run)])==np.array(y_run))[0])/len(y_run)
stable_blocks1 = stable_blocks1[200:,:,:]
stable_blocks1 = np.concatenate((stable_blocks1,np.zeros((200,n_channels,n_samples_fs100))),axis=0)
s_begin = 800+b*400
offset = 400 # Indexing offset for EEG data and y
stable_blocks0 = e[s_begin:s_begin+200,:,:]
y_run = np.concatenate((y_run[200:],y[s_begin:s_begin+200]))
# Test accuracy of RT epochs
for t in range(200):
print('Testing epoch number: ',c)
epoch = e[c,:,:]
epoch = preproc1epoch(epoch,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch+epoch_prev)/2
pred_prob_test[c_test,:],y_pred_test1[c_test] = testEpoch(clf,epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test,0] = np.min([np.max([pred_prob_test[c_test,0]+offset_pred,0]),1])
pred_prob_test_corr[c_test,1] = np.min([np.max([pred_prob_test[c_test,1]-offset_pred,0]),1])
clf_output = pred_prob_test_corr[c_test,int(y[c])]-pred_prob_test_corr[c_test,int(y[c]-1)]
clf_output_test[c_test] = clf_output # Save classifier output for correlation checks
alpha_test[c_test] = sigmoid(clf_output) # Convert corrected classifier output to an alpha value using a sigmoid transfer function
epoch_prev = epoch
epochs_fb[t,:,:] = epoch
c += 1
c_test += 1
above_chance_offline = len(np.where((np.array(alpha_test[:c_test])>0.5))[0])/len(alpha_test[:c_test])
print('Above chance alpha (corrected): ' + str(above_chance_offline))
score = metrics.accuracy_score(y_test_feedback[:c_test], y_pred_test1[:c_test])
print('Accuracy score (uncorrected): ' + str(score))
# Test score per run
a_per_run = np.zeros((n_it))
score_per_run = np.zeros((n_it))
j = 0
for run in range(n_it):
score_per_run[run] = metrics.accuracy_score(y_test_feedback[run*200:(run+1)*200], y_pred_test1[run*200:(run+1)*200])
a_per_run[run] = (len(np.where((np.array(alpha_test[j:j+200])>0.5))[0])/200)
j += 200
print('Alpha, corrected, above chance per run: ' + str(a_per_run))
print('Score, uncorrected, per run: ' + str(score_per_run))
d['RT_train_acc'] = train_acc
d['RT_test_acc_corr'] = above_chance_offline
d['RT_test_acc_corr_run'] = a_per_run
d['RT_test_acc_uncorr'] = score
d['RT_test_acc_uncorr_run'] = score_per_run
#%% Analyze RT session block-wise
d['ALPHA_test'] = alpha_test
d['CLFO_test'] = clf_output_test
#%% Compare RT alpha with RT offline alpha
alphan = np.asarray(alpha)
alpha_corr = (np.corrcoef(alphan[1:], alpha_test[:999]))[0][1] # Shifted with one
d['ALPHA_correlation'] = alpha_corr
if alpha_corr >= 0.98:
d['GROUP'] = 1 # 1 for NF group
if alpha_corr < 0.98:
d['GROUP'] = 0 # 0 for control group
# Classifier output correlation has to be checked offline, because clf_output values are computed offline (RT pipeline)
#np.corrcoef(clf_output_test, clf_output_test13)
#plt.plot(alphan)
#plt.plot(alpha_test)
#
## Compare alphas from file and offline classification
#plt.plot(np.arange(999),alphan[1:]) #blue, en foran
#plt.plot(np.arange(1000),alpha_test)
#
## Run 1
#plt.plot(np.arange(200),alphan[1:201]) #starts at 0.5
#plt.plot(np.arange(200),alpha_test[0:200]) #matches
#%% Confusion matrices
n_it_trials = n_it*200
correct = (y_test_feedback[:n_it_trials]==y_pred_test1[:n_it_trials])
alpha_test_c = np.copy(alpha_test)
alpha_test_c[alpha_test_c > 0.5] = True # 1 is a correctly predicted
alpha_test_c[alpha_test_c < 0.5] = False
alpha_predcat = np.argmax(pred_prob_test_corr,axis=1)
conf_uncorr = confusion_matrix(y_test_feedback[:n_it_trials],y_pred_test1[:n_it_trials])
conf_corr = confusion_matrix(y_test_feedback[:n_it_trials],alpha_predcat[:n_it_trials])
# Separate into scenes and faces accuracy
scene_acc = conf_corr[0,0]/(conf_corr[0,0]+conf_corr[0,1])
face_acc = conf_corr[1,1]/(conf_corr[1,0]+conf_corr[1,1])
d['RT_correct_NFtest_pred'] = correct
d['RT_conf_corr'] = conf_corr
d['RT_conf_uncorr'] = conf_uncorr
d['RT_scene_acc'] = scene_acc
d['RT_face_acc'] = face_acc
# Training confusion matrices
conf_train = []
for b in range(n_it):
conf_train.append(confusion_matrix(Y_run[b,:], Y_train[b,:]))
d['RT_conf_train'] = conf_train
#%% Training on stable blocks only - leave one block out CV
offset_pred_lst = []
c_test = 0
no_sb = 8+4*n_it # Number stable blocks
block_len = 50
pred_prob_test = np.zeros((no_sb*block_len,2)) # Prediction probability test. Block length of 50 trials
pred_prob_test_corr = np.zeros((no_sb*block_len,2)) # Prediction probability test, corrected for bias
alpha_test = np.zeros((no_sb*block_len)) # Alpha values for stable blocks
stable_blocks_fbrun = np.concatenate([e[400+n*400:600+n*400] for n in range(n_it)]) # Stable blocks feedback run
y_stable_blocks_fbrun = np.concatenate([y[400+n*400:600+n*400] for n in range(n_it)])
stable_blocks = np.concatenate((e[:400,:,:], stable_blocks_fbrun))
y_stable_blocks = np.concatenate((y[:400], y_stable_blocks_fbrun))
y_pred = np.zeros(no_sb*block_len)
for sb in range(no_sb):
val_indices = range(sb*block_len,(sb+1)*block_len) # Validation block index
stable_blocks_val = stable_blocks[val_indices]
y_val = y_stable_blocks[val_indices]
stable_blocks_train = np.delete(stable_blocks, val_indices, axis=0)
y_train = np.delete(y_stable_blocks, val_indices)
stable_blocks_train_prep = np.zeros((len(y_train),23,n_samples_fs100))
for t in range(stable_blocks_train.shape[0]):
epoch = stable_blocks_train[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
stable_blocks_train_prep[t,:,:] = epoch
projs1,stable_blocksSSP_train = applySSP(stable_blocks_train_prep,info_fs100,threshold=threshold)
# Average after SSP correction
stable_blocksSSP_train = average_stable(stable_blocksSSP_train)
clf,offset_pred = trainLogReg_cross_offline(stable_blocksSSP_train,y_train) #cur. in EEG_classification
offset_pred = np.min([np.max([offset_pred,-0.25]),0.25])/2
offset_pred_lst.append(offset_pred)
# Test epochs in validation block. Preprocessing and testing epoch-wise
for t in range(block_len):
epoch = stable_blocks_val[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch+epoch_prev)/2
pred_prob_test[c_test,:],y_pred[c_test] = testEpoch(clf,epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test,0] = np.min([np.max([pred_prob_test[c_test,0]+offset_pred,0]),1])
pred_prob_test_corr[c_test,1] = np.min([np.max([pred_prob_test[c_test,1]-offset_pred,0]),1])
clf_output = pred_prob_test_corr[c_test,int(y_val[t])]-pred_prob_test_corr[c_test,int(y_val[t]-1)]
alpha_test[c_test] = sigmoid(clf_output)
epoch_prev = epoch
c_test += 1
print('No c_test: ' + str(c_test) + 'out of ' + str(no_sb*block_len))
above_chance_train = len(np.where((np.array(alpha_test[:c_test])>0.5))[0])/len(alpha_test[:c_test])
print('Above chance alpha train (corrected): ' + str(above_chance_train))
score = metrics.accuracy_score(y_stable_blocks, y_pred)
d['LOBO_stable_train_offsets'] = offset_pred_lst
d['LOBO_stable_train_acc_corr'] = above_chance_train
d['LOBO_stable_train_acc_uncorr'] = score
#%% Extract data for MNE plots
# Perform preprocessing and SSP on all the stable blocks.
stable_blocks_plot = np.zeros((len(y_stable_blocks),n_channels,n_samples_fs100))
for t in range(stable_blocks.shape[0]):
epoch = stable_blocks[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
stable_blocks_plot[t,:,:] = epoch
projs1,stable_blocksSSP_plot,p_variance = applySSP_forplot(stable_blocks_plot,info_fs100,threshold=threshold,add_events=y_stable_blocks)
d['MNE_stable_blocks_SSP'] = stable_blocksSSP_plot
d['MNE_stable_blocks_SSP_projvariance'] = p_variance
d['MNE_y_stable_blocks'] = y_stable_blocks
#%% Confusion matrices - stable blocks accuracy, LOBO
# y_stable_blocks has the correct y vals, y_pred has the predicted vals. For uncorrected prediction.
# Uncorrected
correct = (y_stable_blocks == y_pred)
conf_train_stable_uncorr = confusion_matrix(y_stable_blocks,y_pred)
# Separate into scenes and faces accuracy
scene_acc_uncorr = conf_train_stable_uncorr[0,0]/(conf_train_stable_uncorr[0,0]+conf_train_stable_uncorr[0,1])
face_acc_uncorr = conf_train_stable_uncorr[1,1]/(conf_train_stable_uncorr[1,0]+conf_train_stable_uncorr[1,1])
# Corrected
alpha_test_c = np.copy(alpha_test)
alpha_test_c[alpha_test_c > 0.5] = True # 1 is a correctly predicted
alpha_test_c[alpha_test_c < 0.5] = False
alpha_predcat = np.argmax(pred_prob_test_corr,axis=1)
conf_train_stable = confusion_matrix(y_stable_blocks,alpha_predcat)
# Separate into scenes and faces accuracy
scene_acc = conf_train_stable[0,0]/(conf_train_stable[0,0]+conf_train_stable[0,1])
face_acc = conf_train_stable[1,1]/(conf_train_stable[1,0]+conf_train_stable[1,1])
d['LOBO_stable_conf_uncorr'] = conf_train_stable_uncorr
d['LOBO_stable_scene_acc_uncorr'] = scene_acc_uncorr
d['LOBO_stable_face_acc_uncorr'] = face_acc_uncorr
d['LOBO_stable_conf_corr'] = conf_train_stable
d['LOBO_stable_scene_acc_corr'] = scene_acc
d['LOBO_stable_face_acc_corr'] = face_acc
#%% Training on stable blocks only - leave one run out CV. The first run is not used as test set, only for training.
offset_pred_lst = []
c_test = 0
no_sb = 8+4*n_it # Number stable blocks
block_len = 50
pred_prob_test = np.zeros((no_sb*block_len,2)) # Prediction probability test. Block length of 50 trials
pred_prob_test_corr = np.zeros((no_sb*block_len,2)) # Prediction probability test, corrected for bias
alpha_test = np.zeros((no_sb*block_len)) # Alpha values for stable blocks
stable_blocks_fbrun = np.concatenate([e[400+n*400:600+n*400] for n in range(n_it)]) # Stable blocks feedback run
y_stable_blocks_fbrun = np.concatenate([y[400+n*400:600+n*400] for n in range(n_it)])
stable_blocks = np.concatenate((e[:400,:,:], stable_blocks_fbrun))
y_stable_blocks = np.concatenate((y[:400], y_stable_blocks_fbrun))
y_pred = np.zeros(no_sb*block_len)
for r in range(n_it): # 5 runs
print('Run no: ',r)
val_indices = range((r+2)*200,((r+2)*200)+200) # Validation block index
stable_blocks_val = stable_blocks[val_indices]
y_val = y_stable_blocks[val_indices]
stable_blocks_train = np.delete(stable_blocks, val_indices, axis=0)
y_train = np.delete(y_stable_blocks, val_indices)
stable_blocks_train_prep = np.zeros((len(y_train),23,n_samples_fs100))
for t in range(stable_blocks_train.shape[0]):
epoch = stable_blocks_train[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
stable_blocks_train_prep[t,:,:] = epoch
projs1,stable_blocksSSP_train = applySSP(stable_blocks_train_prep,info_fs100,threshold=threshold)
# Average after SSP correction
stable_blocksSSP_train = average_stable(stable_blocksSSP_train)
clf,offset_pred = trainLogReg_cross_offline(stable_blocksSSP_train,y_train) #cur. in EEG_classification
offset_pred = np.min([np.max([offset_pred,-0.25]),0.25])/2
offset_pred_lst.append(offset_pred)
# Test epochs in validation run. Preprocessing and testing epoch-wise
for t in range(len(val_indices)):
epoch = stable_blocks_val[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch+epoch_prev)/2
pred_prob_test[c_test,:],y_pred[c_test] = testEpoch(clf,epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test,0] = np.min([np.max([pred_prob_test[c_test,0]+offset_pred,0]),1])
pred_prob_test_corr[c_test,1] = np.min([np.max([pred_prob_test[c_test,1]-offset_pred,0]),1])
clf_output = pred_prob_test_corr[c_test,int(y_val[t])]-pred_prob_test_corr[c_test,int(y_val[t]-1)]
alpha_test[c_test] = sigmoid(clf_output)
epoch_prev = epoch
c_test += 1
print('No c_test: ' + str(c_test) + ' out of ' + str(no_sb*block_len))
above_chance_train = len(np.where((np.array(alpha_test[:c_test])>0.5))[0])/len(alpha_test[:c_test])
print('Above chance alpha train (corrected): ' + str(above_chance_train))
score = metrics.accuracy_score(y_stable_blocks, y_pred)
d['LORO_stable_train_offsets_stable'] = offset_pred_lst
d['LORO_stable_acc_corr'] = above_chance_train
d['LORO_stable_acc_uncorr'] = score
#%% Extract RT epochs (non-averaged) for plots and analysis
stable_blocks0 = e[:600,:,:] # First run + stable in run 2
stable_blocks1 = np.zeros((600,n_channels,n_samples_fs100))
y = np.array([int(x) for x in cat])
y_run = y[:600]
c_test = 0
c = 0
offset = 0
y_pred_test1 = np.zeros(5*200)
y_test_feedback = np.concatenate((y[600:800],y[1000:1200],y[1400:1600],y[1800:2000],y[2200:2400]))
epochs_fb_nonavg = np.zeros((1000,23,n_samples_fs100)) # Epochs feedback
# epochs_fb_avg = np.zeros((1000,23,n_samples_fs100)) # Epochs feedback
for b in range(n_it):
for t in range(stable_blocks0.shape[0]):
epoch = stable_blocks0[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=None,mne_reject=0,reject_ch=reject_ch,flat=None,bad_channels=None,opt_detrend=1)
stable_blocks1[t+offset,:,:] = epoch
c += 1
projs1,stable_blocksSSP1 = applySSP(stable_blocks1,info_fs100,threshold=threshold) # Apply SSP on stable blocks
stable_blocks1 = stable_blocks1[200:,:,:]
stable_blocks1 = np.concatenate((stable_blocks1,np.zeros((200,n_channels,n_samples_fs100))),axis=0)
s_begin = 800+b*400
offset = 400 # Indexing offset for EEG data and y
stable_blocks0 = e[s_begin:s_begin+200,:,:]
y_run = np.concatenate((y_run[200:],y[s_begin:s_begin+200]))
# Append RT epochs
for t in range(200):
print('Epoch number: ',c)
epoch1 = e[c,:,:]
epoch1 = preproc1epoch(epoch1,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
# For averaging epochs:
# if t == 0:
# epoch_avg = epoch1
#
# if t > 0:
# epoch_avg = (epoch1+epoch_prev)/2 # Checked again 17 April that this is legit
#
# epoch_prev = epoch_avg
#
# epochs_fb_avg[c_test,:,:] = epoch_avg
epochs_fb_nonavg[c_test,:,:] = epoch1
c += 1
c_test += 1
# Create MNE objects
events_list = y_test_feedback
event_id = dict(scene=0, face=1)
n_epochs = len(events_list)
events_list = [int(i) for i in events_list]
events = np.c_[np.arange(n_epochs), np.zeros(n_epochs, int),events_list]
eRT_nonavg = mne.EpochsArray(epochs_fb_nonavg, info=info_fs100, events=events,event_id=event_id,tmin=-0.1,baseline=None)
# eRT_avg = mne.EpochsArray(epochs_fb_avg, info=info_fs100, events=events,event_id=event_id,tmin=-0.1,baseline=None)
d['MNE_RT_epochs_fb_nonavg'] = eRT_nonavg
# d['MNE_RT_epochs_fb_avg'] = eRT_avg
d['MNE_y_test_feedback'] = y_test_feedback
if plot_MNE == True:
# Creating a dict of lists: Condition 0 and condition 1 with evoked arrays.
evoked_array_c0 = []
evoked_array_c1 = []
eRT_get = eRT_nonavg.get_data()
for idx,cat in enumerate(events_list):
if cat == 0:
evoked_array_c0.append(mne.EvokedArray(eRT_get[idx], info_fs100,tmin=-0.1,comment=cat)) # Scenes 0
print
if cat == 1:
evoked_array_c1.append(mne.EvokedArray(eRT_get[idx], info_fs100,tmin=-0.1,comment=cat)) # Faces 1
e_dict={}
e_dict['0'] = evoked_array_c0
e_dict['1'] = evoked_array_c1
#colors = 'red', 'blue'
#mne.viz.plot_compare_evokeds(e_dict,ci=0.95,picks=[7],colors=colors)
#%% Save pckl file
pkl_arr = [d]
print('Finished running test and train analyses for subject: ' + str(subjID))
# PICKLE TIME
fname = '08May2_subj_'+str(subjID)+'.pkl'
with open(fname, 'wb') as fout:
pickle.dump(pkl_arr, fout)
epochs0_idx = np.where(ss == 0)[0]
if len(epochs0_idx):
rep = 0
while rep < 30:
print('WARNING missing ' + str(len(epochs0_idx)) +
' stable epochs\n')
rep += 1
epochs_non0_avg = np.mean(np.delete(stable_blocks,
epochs0_idx,
axis=0),
axis=0)
stable_blocks[epochs0_idx] = epochs_non0_avg
# Compute SSP projectors based on stable blocks (training EEG data)
projs, stable_blocksSSP = applySSP(stable_blocks,
info_fs100,
threshold=threshold)
# Average two consecutive trials over a moving window
stable_blocksSSP = average_stable(stable_blocksSSP)
print('Training classifier on SSP corrected EEG data')
# Train classifier for other runs than the first run and estimate classifier bias for offset correction
if n_run > 0:
print('Number of run is above 1')
# Test if neurofeedback epochs are missing
ss_fb = np.sum(np.sum(np.abs(epochs_fb), axis=2), axis=1)
epochs0_idx = np.where(ss_fb == 0)[0]
if len(epochs0_idx):
rep = 0
def analyzeOffline(subjID):
'''
'''
# Initialize conditions for preprocessing of epochs (preproc1epoch from EEG_analysis_RT)
reject_ch = 1 # Rejection of nine predefined channels
reject = None # Rejection of channels, either manually defined or based on MNE analysis
mne_reject = 0
flat = None # Input for MNE rejection
bad_channels = None # Input for manual rejection of channels
opt_detrend = 1 # Temporal EEG detrending (linear)
if reject_ch == 1:
n_channels = 23
if reject_ch == 0:
n_channels = 32
# Initialize conditions for SSP rejection (applySSP from EEG_analysis_RT)
threshold = 0.1 # SSP projection variance threshold
# Initialize dictionary for saving outputs
d = {}
d['subjID'] = subjID
plot_MNE = 0
EEGfile, markerFile, idxFile, alphaFile, n_it = findSubjectFiles(subjID,manual_n_it=None)
#%% Test RT alpha per run
alpha,marker = extractAlpha(alphaFile)
above_chance = len(np.where((np.array(alpha)>0.5))[0])/len(alpha)
alpha_per_run = np.zeros((n_it))
j = 0
for ii in range(n_it):
alpha_per_run[ii] = len(np.where((np.array(alpha[j:j+200])>0.5))[0])/200
j += 200
d['ALPHA_fromfile_overall'] = above_chance
d['ALPHA_fromfile_run'] = alpha_per_run
#%% Extract epochs from EEG data
prefilter = 0
n_samples_fs500 = 450 # Number of samples to extract for each epoch, sampling frequency 500
n_samples_fs100 = int(450/5) # Number of samples, sampling frequency 100 (resampled)
e = extractEpochs_tmin(EEGfile,markerFile,prefilter=prefilter,marker1=0,n_samples=n_samples_fs500)
cat = extractCat(idxFile,exp_type='fused')
# MNE info files
info_fs500 = create_info_mne(reject_ch=0,sfreq=500)
info_fs100 = create_info_mne(reject_ch=1,sfreq=100)
#%% Run NF RT offline analysis
stable_blocks0 = e[:600,:,:] # Fi+st run
stable_blocks1 = np.zeros((600,n_channels,n_samples_fs100))
y = np.array([int(x) for x in cat])
y_run = y[:600]
pred_prob_train = np.zeros((n_it*600,2))
pred_prob_test = np.zeros((n_it*200,2)) # Prediction probability
pred_prob_test_corr = np.zeros((n_it*200,2)) # Prediction probability, corrected for classifier bias
alpha_test = np.zeros((n_it*200))
clf_output_test = np.zeros((n_it*200))
train_acc = np.zeros(n_it)
c_test = 0
c = 0
offset = 0
y_pred_test1 = np.zeros(5*200)
y_test_feedback = np.concatenate((y[600:800],y[1000:1200],y[1400:1600],y[1800:2000],y[2200:2400]))
Y_train = np.zeros((n_it,600))
Y_run = np.zeros((n_it,600))
val = 1 # Offset validation
offset_pred = 0 # Offset prediction
offset_Pred = []
# Score = np.zeros((n_it,3))
epochs_fb = np.zeros((200,23,n_samples_fs100)) # Epochs feedback
for b in range(n_it):
for t in range(stable_blocks0.shape[0]):
epoch = stable_blocks0[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
stable_blocks1[t+offset,:,:] = epoch
c += 1
projs1,stable_blocksSSP1 = applySSP(stable_blocks1,info_fs100,threshold=threshold) # Apply SSP on stable blocks
# Average training blocks after SSP
stable_blocksSSP1 = average_stable(stable_blocksSSP1)
if val == 1: # Test for offset classifier bias
if b > 0: # Runs after first run
stable_blocksSSP1_append = np.append(stable_blocksSSP1,epochs_fb,axis=0)
fb_start = (b-1)*200
y_run_append = np.append(y_run,y_test_feedback[fb_start:fb_start+200],axis=0)
clf,offset_pred = trainLogReg_cross2(stable_blocksSSP1_append,y_run_append)
else:
clf,offset_pred = trainLogReg_cross(stable_blocksSSP1,y_run) # First run
offset_Pred.append(offset_pred)
#Score[b,:]=score
print('Offset estimated to '+str(offset_pred))
else:
clf = trainLogReg(stable_blocksSSP1,y_run)
Y_run[b,:]=y_run
y_pred_train1 = np.zeros(len(y_run))
offset_pred = np.min([np.max([offset_pred,-0.25]),0.25])/2 # Limit offset correction to abs(0.125)
# Training accuracy based on the RT classifier
for t in range(len(y_run)):
epoch_t = stable_blocksSSP1[t,:,:]
pred_prob_train[t,:],y_pred_train1[t] = testEpoch(clf,epoch_t)
Y_train[b,:]=y_pred_train1
train_acc[b] = len(np.where(np.array(y_pred_train1[0:len(y_run)])==np.array(y_run))[0])/len(y_run)
stable_blocks1 = stable_blocks1[200:,:,:]
stable_blocks1 = np.concatenate((stable_blocks1,np.zeros((200,n_channels,n_samples_fs100))),axis=0)
s_begin = 800+b*400
offset = 400 # Indexing offset for EEG data and y
stable_blocks0 = e[s_begin:s_begin+200,:,:]
y_run = np.concatenate((y_run[200:],y[s_begin:s_begin+200]))
# Test accuracy of RT epochs
for t in range(200):
print('Testing epoch number: ',c)
epoch = e[c,:,:]
epoch = preproc1epoch(epoch,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch+epoch_prev)/2
pred_prob_test[c_test,:],y_pred_test1[c_test] = testEpoch(clf,epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test,0] = np.min([np.max([pred_prob_test[c_test,0]+offset_pred,0]),1])
pred_prob_test_corr[c_test,1] = np.min([np.max([pred_prob_test[c_test,1]-offset_pred,0]),1])
clf_output = pred_prob_test_corr[c_test,int(y[c])]-pred_prob_test_corr[c_test,int(y[c]-1)]
clf_output_test[c_test] = clf_output # Save classifier output for correlation checks
alpha_test[c_test] = sigmoid(clf_output) # Convert corrected classifier output to an alpha value using a sigmoid transfer function
epoch_prev = epoch
epochs_fb[t,:,:] = epoch
c += 1
c_test += 1
above_chance_offline = len(np.where((np.array(alpha_test[:c_test])>0.5))[0])/len(alpha_test[:c_test])
print('Above chance alpha (corrected): ' + str(above_chance_offline))
score = metrics.accuracy_score(y_test_feedback[:c_test], y_pred_test1[:c_test])
print('Accuracy score (uncorrected): ' + str(score))
# Test score per run
a_per_run = np.zeros((n_it))
score_per_run = np.zeros((n_it))
j = 0
for run in range(n_it):
score_per_run[run] = metrics.accuracy_score(y_test_feedback[run*200:(run+1)*200], y_pred_test1[run*200:(run+1)*200])
a_per_run[run] = (len(np.where((np.array(alpha_test[j:j+200])>0.5))[0])/200)
j += 200
print('Alpha, corrected, above chance per run: ' + str(a_per_run))
print('Score, uncorrected, per run: ' + str(score_per_run))
d['RT_train_acc'] = train_acc
d['RT_test_acc_corr'] = above_chance_offline
d['RT_test_acc_corr_run'] = a_per_run
d['RT_test_acc_uncorr'] = score
d['RT_test_acc_uncorr_run'] = score_per_run
#%% Compare RT alpha with RT offline alpha
alphan = np.asarray(alpha)
alpha_corr = (np.corrcoef(alphan[1:], alpha_test[:999]))[0][1] # Shifted with one
d['ALPHA_correlation'] = alpha_corr
if alpha_corr >= 0.98:
d['GROUP'] = 1 # 1 for NF group
if alpha_corr < 0.98:
d['GROUP'] = 0 # 0 for control group
# Classifier output correlation has to be checked offline, because clf_output values are computed offline (RT pipeline)
#np.corrcoef(clf_output_test, clf_output_test13)
#plt.plot(alphan)
#plt.plot(alpha_test)
#
## Compare alphas from file and offline classification
#plt.plot(np.arange(999),alphan[1:]) #blue, en foran
#plt.plot(np.arange(1000),alpha_test)
#
## Run 1
#plt.plot(np.arange(200),alphan[1:201]) #starts at 0.5
#plt.plot(np.arange(200),alpha_test[0:200]) #matches
#%% Confusion matrices
n_it_trials = n_it*200
correct = (y_test_feedback[:n_it_trials]==y_pred_test1[:n_it_trials])
alpha_test_c = np.copy(alpha_test)
alpha_test_c[alpha_test_c > 0.5] = True # 1 is a correctly predicted
alpha_test_c[alpha_test_c < 0.5] = False
alpha_predcat = np.argmax(pred_prob_test_corr,axis=1)
conf_uncorr = confusion_matrix(y_test_feedback[:n_it_trials],y_pred_test1[:n_it_trials])
conf_corr = confusion_matrix(y_test_feedback[:n_it_trials],alpha_predcat[:n_it_trials])
# Separate into scenes and faces accuracy
scene_acc = conf_corr[0,0]/(conf_corr[0,0]+conf_corr[0,1])
face_acc = conf_corr[1,1]/(conf_corr[1,0]+conf_corr[1,1])
d['RT_correct_NFtest_pred'] = correct
d['RT_conf_corr'] = conf_corr
d['RT_conf_uncorr'] = conf_uncorr
d['RT_scene_acc'] = scene_acc
d['RT_face_acc'] = face_acc
# Training confusion matrices
conf_train = []
for b in range(n_it):
conf_train.append(confusion_matrix(Y_run[b,:], Y_train[b,:]))
d['RT_conf_train'] = conf_train
#%% Training on stable blocks only - leave one block out CV
offset_pred_lst = []
c_test = 0
no_sb = 8+4*n_it # Number stable blocks
block_len = 50
pred_prob_test = np.zeros((no_sb*block_len,2)) # Prediction probability test. Block length of 50 trials
pred_prob_test_corr = np.zeros((no_sb*block_len,2)) # Prediction probability test, corrected for bias
alpha_test = np.zeros((no_sb*block_len)) # Alpha values for stable blocks
stable_blocks_fbrun = np.concatenate([e[400+n*400:600+n*400] for n in range(n_it)]) # Stable blocks feedback run
y_stable_blocks_fbrun = np.concatenate([y[400+n*400:600+n*400] for n in range(n_it)])
stable_blocks = np.concatenate((e[:400,:,:], stable_blocks_fbrun))
y_stable_blocks = np.concatenate((y[:400], y_stable_blocks_fbrun))
y_pred = np.zeros(no_sb*block_len)
for sb in range(no_sb):
val_indices = range(sb*block_len,(sb+1)*block_len) # Validation block index
stable_blocks_val = stable_blocks[val_indices]
y_val = y_stable_blocks[val_indices]
stable_blocks_train = np.delete(stable_blocks, val_indices, axis=0)
y_train = np.delete(y_stable_blocks, val_indices)
stable_blocks_train_prep = np.zeros((len(y_train),23,n_samples_fs100))
for t in range(stable_blocks_train.shape[0]):
epoch = stable_blocks_train[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
stable_blocks_train_prep[t,:,:] = epoch
projs1,stable_blocksSSP_train = applySSP(stable_blocks_train_prep,info_fs100,threshold=threshold)
# Average after SSP correction
stable_blocksSSP_train = average_stable(stable_blocksSSP_train)
clf,offset_pred = trainLogReg_cross_offline(stable_blocksSSP_train,y_train) #cur. in EEG_classification
offset_pred = np.min([np.max([offset_pred,-0.25]),0.25])/2
offset_pred_lst.append(offset_pred)
# Test epochs in validation block. Preprocessing and testing epoch-wise
for t in range(block_len):
epoch = stable_blocks_val[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch+epoch_prev)/2
pred_prob_test[c_test,:],y_pred[c_test] = testEpoch(clf,epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test,0] = np.min([np.max([pred_prob_test[c_test,0]+offset_pred,0]),1])
pred_prob_test_corr[c_test,1] = np.min([np.max([pred_prob_test[c_test,1]-offset_pred,0]),1])
clf_output = pred_prob_test_corr[c_test,int(y_val[t])]-pred_prob_test_corr[c_test,int(y_val[t]-1)]
alpha_test[c_test] = sigmoid(clf_output)
epoch_prev = epoch
c_test += 1
print('No c_test: ' + str(c_test) + 'out of ' + str(no_sb*block_len))
above_chance_train = len(np.where((np.array(alpha_test[:c_test])>0.5))[0])/len(alpha_test[:c_test])
print('Above chance alpha train (corrected): ' + str(above_chance_train))
score = metrics.accuracy_score(y_stable_blocks, y_pred)
d['LOBO_stable_train_offsets'] = offset_pred_lst
d['LOBO_stable_train_acc_corr'] = above_chance_train
d['LOBO_stable_train_acc_uncorr'] = score
#%% Extract data for MNE plots
# Perform preprocessing and SSP on all the stable blocks.
stable_blocks_plot = np.zeros((len(y_stable_blocks),n_channels,n_samples_fs100))
for t in range(stable_blocks.shape[0]):
epoch = stable_blocks[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
stable_blocks_plot[t,:,:] = epoch
projs1,stable_blocksSSP_plot,p_variance = applySSP_forplot(stable_blocks_plot,info_fs100,threshold=threshold,add_events=y_stable_blocks)
d['MNE_stable_blocks_SSP'] = stable_blocksSSP_plot
d['MNE_stable_blocks_SSP_projvariance'] = p_variance
d['MNE_y_stable_blocks'] = y_stable_blocks
if plot_MNE == True:
print('Making a lot of MNE plots!')
#%% Adding events manually to epochs *also implemented in applySSP_forplot*
stable_blocksSSP_get = stable_blocksSSP_plot.get_data()
events_list = y_stable_blocks
event_id = dict(scene=0, face=1)
n_epochs = len(events_list)
events_list = [int(i) for i in events_list]
events = np.c_[np.arange(n_epochs), np.zeros(n_epochs, int),events_list]
epochs_events = mne.EpochsArray(stable_blocksSSP_get, info_fs100, events=events, tmin=-0.1, event_id=event_id,baseline=None)
epochs_events['face'].average().plot()
epochs_events['scene'].average().plot()
#%%
# Overall average plot, all channels
stable_blocksSSP_plot.average().plot(spatial_colors=True, time_unit='s')#,picks=[7])
# Plot based on categories
stable_blocksSSP_plot['face'].average().plot(spatial_colors=True, time_unit='s')#,picks=[7])
stable_blocksSSP_plot['scene'].average().plot(spatial_colors=True, time_unit='s')
# Plot of the SSP projectors
stable_blocksSSP_plot.average().plot_projs_topomap()
# Consider adding p_variance values to the plot.
# Plot the topomap of the power spectral density across epochs.
stable_blocksSSP_plot.plot_psd_topomap(proj=True)
stable_blocksSSP_plot['face'].plot_psd_topomap(proj=True)
stable_blocksSSP_plot['scene'].plot_psd_topomap(proj=True)
# Plot topomap (possibiity of adding specific times)
stable_blocksSSP_plot.average().plot_topomap(proj=True)
stable_blocksSSP_plot.average().plot_topomap(proj=True,times=np.linspace(0.05, 0.15, 5))
# Plot joint topomap and evoked ERP
stable_blocksSSP_plot.average().plot_joint()
# If manually adding a sensor plot
stable_blocksSSP_plot.plot_sensors(show_names=True)
# Noise covariance plot - not really sure what to make of this (yet)
noise_cov = mne.compute_covariance(stable_blocksSSP_plot)
fig = mne.viz.plot_cov(noise_cov, stable_blocksSSP_plot.info)
# Generate list of evoked objects from condition names
evokeds = [stable_blocksSSP_plot[name].average() for name in ('scene','face')]
colors = 'blue', 'red'
title = 'Subject \nscene vs face'
# Plot evoked across all channels, comparing two categories
mne.viz.plot_evoked_topo(evokeds, color=colors, title=title, background_color='w')
# Compare two categories
mne.viz.plot_compare_evokeds(evokeds,title=title,show_sensors=True,cmap='viridis')#,ci=True)
# When multiple channels are passed, this function combines them all, to get one time course for each condition.
mne.viz.plot_compare_evokeds(evokeds,title=title,show_sensors=True,cmap='viridis',ci=True,picks=[7])
# Make animation
fig,anim = evokeds[0].animate_topomap(times=np.linspace(0.00, 0.79, 100),butterfly=True)
# Save animation
fig,anim = evokeds[0].animate_topomap(times=np.linspace(0.00, 0.79, 50),frame_rate=10,blit=False)
anim.save('Brainmation.gif', writer='imagemagick', fps=10)
# Sort epochs based on categories
sorted_epochsarray = [stable_blocksSSP_plot[name] for name in ('scene','face')]
# Plot image
stable_blocksSSP_plot.plot_image()
# Appending all entries in the overall epochsarray as single evoked arrays of shape (n_channels, n_times)
g2 = stable_blocksSSP_plot.get_data()
evoked_array = [mne.EvokedArray(entry, info_fs100,tmin=-0.1) for entry in g2]
# Appending all entries in the overall epochsarray as single evoked arrays of shape (n_channels, n_times) - category info added as comments
evoked_array2 = []
for idx,cat in enumerate(events_list):
evoked_array2.append(mne.EvokedArray(g2[idx], info_fs100,tmin=-0.1,comment=cat))
mne.viz.plot_compare_evokeds(evoked_array2[:10]) # Plotting all the individual evoked arrays (up to 10)
# Creating a dict of lists: Condition 0 and condition 1 with evoked arrays.
evoked_array_c0 = []
evoked_array_c1 = []
for idx,cat in enumerate(events_list):
if cat == 0:
evoked_array_c0.append(mne.EvokedArray(g2[idx], info_fs100,tmin=-0.1,comment=cat)) # Scenes 0
if cat == 1:
evoked_array_c1.append(mne.EvokedArray(g2[idx], info_fs100,tmin=-0.1,comment=cat)) # Faces 1
e_dict={}
e_dict['0'] = evoked_array_c0
e_dict['1'] = evoked_array_c1
# Could create these e_dicts for several people, and plot the means. Or create an e_dict with the evokeds for each person, and make the "overall" mean with individual evoked means across subjects.
mne.viz.plot_compare_evokeds(e_dict,ci=0.4,picks=[7])#,title=title,show_sensors=True,cmap='viridis',ci=True)
mne.viz.plot_compare_evokeds(e_dict,ci=0.8,picks=[7])#,title=title,show_sensors=True,cmap='viridis',ci=True)
# Comparing
mne.viz.plot_compare_evokeds(evokeds,title=title,show_sensors=True,picks=[7],ci=False)
mne.viz.plot_compare_evokeds(e_dict,title=title,show_sensors=True,picks=[7],ci=True)#,,ci=True)
# Based on categories, the correct epochs are added to the evoked_array_c0 and c1
#%% Manually check whether the MNE events correspond to the actual categories
g1 = stable_blocksSSP_plot.get_data()
g3 = g1[y_stable_blocks==True] # Faces = 1
g4 = g1[y_stable_blocks==False] # Scenes = 0
g3a = np.mean(g3,axis=0)
g4a = np.mean(g4,axis=0)
plt.figure(4)
plt.plot(g3a.T[:,7]) # Corresponds to: stable_blocksSSP_plot['face'].average().plot(spatial_colors=True, time_unit='s')
plt.plot(g4a.T[:,7])
epochsg3 = mne.EpochsArray(g3, info=info_fs100)
epochsg4 = mne.EpochsArray(g4, info=info_fs100)
plt.figure(6)
epochsg3.average().plot(spatial_colors=True, time_unit='s',picks=[7]) # Corresponds to: stable_blocksSSP_plot['face'].average().plot(spatial_colors=True, time_unit='s',picks=[7])
epochsg4.average().plot(spatial_colors=True, time_unit='s',picks=[7])
# Plot topomap
a3 = epochsg3.average()
a3.plot_topomap(times=np.linspace(0.05, 0.15, 5))
# Plot of evoked ERP and topomaps, in one plot!
a3.plot_joint()
# Control the y axis
a3.plot(ylim=dict(eeg=[-2000000, 2000000]))
a3.plot(ylim=dict(eeg=[-2000000, 2000000]))
#%% Confusion matrices - stable blocks accuracy, LOBO
# y_stable_blocks has the correct y vals, y_pred has the predicted vals. For uncorrected prediction.
# Uncorrected
correct = (y_stable_blocks == y_pred)
conf_train_stable_uncorr = confusion_matrix(y_stable_blocks,y_pred)
# Separate into scenes and faces accuracy
scene_acc_uncorr = conf_train_stable_uncorr[0,0]/(conf_train_stable_uncorr[0,0]+conf_train_stable_uncorr[0,1])
face_acc_uncorr = conf_train_stable_uncorr[1,1]/(conf_train_stable_uncorr[1,0]+conf_train_stable_uncorr[1,1])
# Corrected
alpha_test_c = np.copy(alpha_test)
alpha_test_c[alpha_test_c > 0.5] = True # 1 is a correctly predicted
alpha_test_c[alpha_test_c < 0.5] = False
alpha_predcat = np.argmax(pred_prob_test_corr,axis=1)
conf_train_stable = confusion_matrix(y_stable_blocks,alpha_predcat)
# Separate into scenes and faces accuracy
scene_acc = conf_train_stable[0,0]/(conf_train_stable[0,0]+conf_train_stable[0,1])
face_acc = conf_train_stable[1,1]/(conf_train_stable[1,0]+conf_train_stable[1,1])
d['LOBO_stable_conf_uncorr'] = conf_train_stable_uncorr
d['LOBO_stable_scene_acc_uncorr'] = scene_acc_uncorr
d['LOBO_stable_face_acc_uncorr'] = face_acc_uncorr
d['LOBO_stable_conf_corr'] = conf_train_stable
d['LOBO_stable_scene_acc_corr'] = scene_acc
d['LOBO_stable_face_acc_corr'] = face_acc
#%% Training accuracy, training on stable and NF - leave one block out CV. Accuracy can be based on either stable+NF, only stable or only NF blocks
offset_pred_lst = []
c_test = 0
no_b = 8+8*n_it # Number blocks total
pred_prob_test = np.zeros((no_b*block_len,2))
pred_prob_test_corr = np.zeros((no_b*block_len,2))
alpha_test = np.zeros((no_b*block_len))
y_pred = np.zeros(no_b*block_len)
for b in range(no_b):
val_indices = range(b*block_len,(b+1)*block_len)
blocks_val = e[val_indices]
y_val = y[val_indices]
blocks_train = np.delete(e, val_indices,axis=0)
y_train = np.delete(y, val_indices)
blocks_train_prep = np.zeros((len(y_train),23,n_samples_fs100))
for t in range(blocks_train_prep.shape[0]):
epoch = blocks_train[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
blocks_train_prep[t,:,:] = epoch
projs1,blocksSSP_train = applySSP(blocks_train_prep,info_fs100,threshold=threshold)
# Average after SSP correction
blocksSSP_train = average_stable(blocksSSP_train)
clf,offset_pred = trainLogReg_cross_offline(blocksSSP_train,y_train) #cur. in EEG_classification
offset_pred = np.min([np.max([offset_pred,-0.25]),0.25])/2
offset_pred_lst.append(offset_pred)
# Test epochs in left out block
for t in range(block_len):
epoch = blocks_val[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch+epoch_prev)/2
pred_prob_test[c_test,:],y_pred[c_test] = testEpoch(clf,epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test,0] = np.min([np.max([pred_prob_test[c_test,0]+offset_pred,0]),1])
pred_prob_test_corr[c_test,1] = np.min([np.max([pred_prob_test[c_test,1]-offset_pred,0]),1])
clf_output = pred_prob_test_corr[c_test,int(y_val[t])]-pred_prob_test_corr[c_test,int(y_val[t]-1)]
alpha_test[c_test] = sigmoid(clf_output)
epoch_prev = epoch
c_test += 1
print('No c_test: ' + str(c_test) + ' out of ' + str(no_b*block_len))
above_chance_train = len(np.where((np.array(alpha_test[:c_test])>0.5))[0])/len(alpha_test[:c_test])
print('Above chance alpha train (corrected) on both stable and NF: ' + str(above_chance_train))
# Separate in stable only, and stable + NF
e_mock = np.arange((8+n_it*8)*block_len)
stable_blocks_fbrun = np.concatenate([e_mock[400+n*400:600+n*400] for n in range(n_it)]) # Stable blocks feedback run
stable_blocks_idx = np.concatenate((e_mock[:400],stable_blocks_fbrun))
a = alpha_test[stable_blocks_idx]
above_chance_stable = len(np.where((np.array(a)>0.5))[0])/len(a)
print('Above chance alpha train (corrected) on stable blocks: ' + str(above_chance_stable))
#nf_blocks_idx = np.concatenate([e_mock[600+n*400:800+n*400] for n in range(n_it)]) # Neurofeedback blocks
#a2 = alpha_test[nf_blocks_idx]
#above_chance_nf = len(np.where((np.array(a2)>0.5))[0])/len(a2)
#print('Above chance alpha train (corrected) on NF blocks: ' + str(above_chance_nf))
d['LOBO_all_train_offsets'] = offset_pred_lst
d['LOBO_all_train_acc'] = above_chance_train # All blocks included
d['LOBO_all_train_acc_stable_test'] = above_chance_stable # Trained on both stable and NF, only tested on stable
#d['train_acc_nf_test'] = above_chance_nf # Trained on both stable and NF, only tested on NF
#%% Confusion matrices - training on stable+NF blocks, testing on stable+NF, stable or NF blocks
# Uncorrected, all (stable + NF)
correct = (y == y_pred)
conf_train_all_uncorr = confusion_matrix(y,y_pred)
# Separate into scenes and faces accuracy
scene_acc_uncorr = conf_train_all_uncorr[0,0]/(conf_train_all_uncorr[0,0]+conf_train_all_uncorr[0,1])
face_acc_uncorr = conf_train_all_uncorr[1,1]/(conf_train_all_uncorr[1,0]+conf_train_all_uncorr[1,1])
# Corrected, all
alpha_test_c = np.copy(alpha_test)
alpha_test_c[alpha_test_c > 0.5] = True # 1 is a correctly predicted
alpha_test_c[alpha_test_c < 0.5] = False
alpha_predcat = np.argmax(pred_prob_test_corr,axis=1)
conf_train_all = confusion_matrix(y,alpha_predcat)
# Separate into scenes and faces accuracy
scene_acc = conf_train_all[0,0]/(conf_train_all[0,0]+conf_train_all[0,1])
face_acc = conf_train_all[1,1]/(conf_train_all[1,0]+conf_train_all[1,1])
d['LOBO_all_conf_uncorr'] = conf_train_all_uncorr
d['LOBO_all_scene_acc_uncorr'] = scene_acc_uncorr
d['LOBO_all_face_acc_uncorr'] = face_acc_uncorr
d['LOBO_all_conf_corr'] = conf_train_all
d['LOBO_all_scene_acc_corr'] = scene_acc
d['LOBO_all_face_acc_corr'] = face_acc
#%% Training on stable blocks only - leave one run out CV
offset_pred_lst = []
c_test = 0
no_sb = 8+4*n_it # Number stable blocks
block_len = 50
pred_prob_test = np.zeros((no_sb*block_len,2)) # Prediction probability test. Block length of 50 trials
pred_prob_test_corr = np.zeros((no_sb*block_len,2)) # Prediction probability test, corrected for bias
alpha_test = np.zeros((no_sb*block_len)) # Alpha values for stable blocks
stable_blocks_fbrun = np.concatenate([e[400+n*400:600+n*400] for n in range(n_it)]) # Stable blocks feedback run
y_stable_blocks_fbrun = np.concatenate([y[400+n*400:600+n*400] for n in range(n_it)])
stable_blocks = np.concatenate((e[:400,:,:], stable_blocks_fbrun))
y_stable_blocks = np.concatenate((y[:400], y_stable_blocks_fbrun))
y_pred = np.zeros(no_sb*block_len)
for r in range(n_it+1): # 6 runs
print('Run no: ',r)
if r == 0: # First run
val_indices = range(0,400) # First run
stable_blocks_val = stable_blocks[val_indices]
y_val = y_stable_blocks[val_indices]
if r > 0:
val_indices = range((r+1)*200,((r+1)*200)+200) # Validation block index
stable_blocks_val = stable_blocks[val_indices]
y_val = y_stable_blocks[val_indices]
stable_blocks_train = np.delete(stable_blocks, val_indices, axis=0)
y_train = np.delete(y_stable_blocks, val_indices)
stable_blocks_train_prep = np.zeros((len(y_train),23,n_samples_fs100))
for t in range(stable_blocks_train.shape[0]):
epoch = stable_blocks_train[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
stable_blocks_train_prep[t,:,:] = epoch
projs1,stable_blocksSSP_train = applySSP(stable_blocks_train_prep,info_fs100,threshold=threshold)
# Average after SSP correction
stable_blocksSSP_train = average_stable(stable_blocksSSP_train)
clf,offset_pred = trainLogReg_cross_offline(stable_blocksSSP_train,y_train) #cur. in EEG_classification
offset_pred = np.min([np.max([offset_pred,-0.25]),0.25])/2
offset_pred_lst.append(offset_pred)
# Test epochs in validation run. Preprocessing and testing epoch-wise
for t in range(len(val_indices)):
epoch = stable_blocks_val[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
if t > 0:
epoch = (epoch+epoch_prev)/2
pred_prob_test[c_test,:],y_pred[c_test] = testEpoch(clf,epoch)
# Correct the prediction bias offset
pred_prob_test_corr[c_test,0] = np.min([np.max([pred_prob_test[c_test,0]+offset_pred,0]),1])
pred_prob_test_corr[c_test,1] = np.min([np.max([pred_prob_test[c_test,1]-offset_pred,0]),1])
clf_output = pred_prob_test_corr[c_test,int(y_val[t])]-pred_prob_test_corr[c_test,int(y_val[t]-1)]
alpha_test[c_test] = sigmoid(clf_output)
epoch_prev = epoch
c_test += 1
print('No c_test: ' + str(c_test) + ' out of ' + str(no_sb*block_len))
above_chance_train = len(np.where((np.array(alpha_test[:c_test])>0.5))[0])/len(alpha_test[:c_test])
print('Above chance alpha train (corrected): ' + str(above_chance_train))
score = metrics.accuracy_score(y_stable_blocks, y_pred)
d['LORO_stable_train_offsets_stable'] = offset_pred_lst
d['LORO_stable_acc_corr'] = above_chance_train
d['LORO_stable_acc_uncorr'] = score
#%% Extract RT epochs (non-averaged) for plots and analysis
stable_blocks0 = e[:600,:,:] # First run
stable_blocks1 = np.zeros((600,n_channels,n_samples_fs100))
y = np.array([int(x) for x in cat])
y_run = y[:600]
c_test = 0
c = 0
offset = 0
y_pred_test1 = np.zeros(5*200)
y_test_feedback = np.concatenate((y[600:800],y[1000:1200],y[1400:1600],y[1800:2000],y[2200:2400]))
epochs_fb_nonavg = np.zeros((1000,23,n_samples_fs100)) # Epochs feedback
# epochs_fb_avg = np.zeros((1000,23,n_samples_fs100)) # Epochs feedback
for b in range(n_it):
for t in range(stable_blocks0.shape[0]):
epoch = stable_blocks0[t,:,:]
epoch = preproc1epoch(epoch,info_fs500,SSP=False,reject=None,mne_reject=0,reject_ch=reject_ch,flat=None,bad_channels=None,opt_detrend=1)
stable_blocks1[t+offset,:,:] = epoch
c += 1
projs1,stable_blocksSSP1 = applySSP(stable_blocks1,info_fs100,threshold=threshold) # Apply SSP on stable blocks
stable_blocks1 = stable_blocks1[200:,:,:]
stable_blocks1 = np.concatenate((stable_blocks1,np.zeros((200,n_channels,n_samples_fs100))),axis=0)
s_begin = 800+b*400
offset = 400 # Indexing offset for EEG data and y
stable_blocks0 = e[s_begin:s_begin+200,:,:]
y_run = np.concatenate((y_run[200:],y[s_begin:s_begin+200]))
# Append RT epochs
for t in range(200):
print('Epoch number: ',c)
epoch1 = e[c,:,:]
epoch1 = preproc1epoch(epoch1,info_fs500,projs=projs1,SSP=True,reject=reject,mne_reject=mne_reject,reject_ch=reject_ch,flat=flat,bad_channels=bad_channels,opt_detrend=opt_detrend)
# For averaging epochs:
# if t == 0:
# epoch_avg = epoch1
#
# if t > 0:
# epoch_avg = (epoch1+epoch_prev)/2 # Checked again 17 April that this is legit
#
# epoch_prev = epoch_avg
#
# epochs_fb_avg[c_test,:,:] = epoch_avg
epochs_fb_nonavg[c_test,:,:] = epoch1
c += 1
c_test += 1
# Create MNE objects
events_list = y_test_feedback
event_id = dict(scene=0, face=1)
n_epochs = len(events_list)
events_list = [int(i) for i in events_list]
events = np.c_[np.arange(n_epochs), np.zeros(n_epochs, int),events_list]
eRT_nonavg = mne.EpochsArray(epochs_fb_nonavg, info=info_fs100, events=events,event_id=event_id,tmin=-0.1,baseline=None)
# eRT_avg = mne.EpochsArray(epochs_fb_avg, info=info_fs100, events=events,event_id=event_id,tmin=-0.1,baseline=None)
d['MNE_RT_epochs_fb_nonavg'] = eRT_nonavg
# d['MNE_RT_epochs_fb_avg'] = eRT_avg
d['MNE_y_test_feedback'] = y_test_feedback
if plot_MNE == True:
# Creating a dict of lists: Condition 0 and condition 1 with evoked arrays.
evoked_array_c0 = []
evoked_array_c1 = []
eRT_get = eRT_nonavg.get_data()
for idx,cat in enumerate(events_list):
if cat == 0:
evoked_array_c0.append(mne.EvokedArray(eRT_get[idx], info_fs100,tmin=-0.1,comment=cat)) # Scenes 0
print
if cat == 1:
evoked_array_c1.append(mne.EvokedArray(eRT_get[idx], info_fs100,tmin=-0.1,comment=cat)) # Faces 1
e_dict={}
e_dict['0'] = evoked_array_c0
e_dict['1'] = evoked_array_c1
#colors = 'red', 'blue'
#mne.viz.plot_compare_evokeds(e_dict,ci=0.95,picks=[7],colors=colors)
#%% Save pckl file
pkl_arr = [d]
print('Finished running test and train analyses for subject: ' + str(subjID))
# PICKLE TIME
fname = '18April_subj_'+str(subjID)+'.pkl'
with open(fname, 'wb') as fout:
pickle.dump(pkl_arr, fout)