def test_OLS():
regress = regression(method=['OLS'],
yrange=[0.0, 100.0],
params=[{
'fit_intercept': True
}],
ransacparams={})
def test_PLS():
regress = regression(method=['PLS'],
yrange=[0.0, 100.0],
params=[{
'n_components': 0,
'scale': False
}],
ransacparams={})
def test_OMP():
regress = regression(method=['OMP'],
yrange=[0.0, 100.0],
params=[{
'fit_intercept': True,
'n_nonzero_coefs': 615,
'CV': True
}],
ransacparams={})
def test_GP():
regress = regression(method=['GP'],
yrange=[0.0, 100.0],
params=[{
'reduce_dim': 'PCA',
'n_components': 0,
'random_start': 1,
'theta0': 1.0,
'thetaL': 0.1,
'thetaU': 100.0
}],
ransacparams={})
def test_KRR():
regress = regression(method=['KRR'],
yrange=[0.0, 100.0],
params=[{
'alpha': 0,
'kernel': 'linear',
'gamma': 'None',
'degree': 3.0,
'coef0': 1.0,
'kernel_params': 'None'
}],
ransacparams={})
def test_Lasso():
regress = regression(method=['Lasso'],
yrange=[0.0, 100.0],
params=[{
'alpha': 1.0,
'fit_intercept': True,
'max_iter': 1000,
'tol': 0.0001,
'positive': False,
'selection': 'random',
'CV': True
}],
ransacparams={})
def test_Ridge():
regress = regression(method=['Ridge'],
yrange=[0.0, 100.0],
params=[{
'alpha': 1.0,
'copy_X': True,
'fit_intercept': True,
'max_iter': 'None',
'normalize': False,
'solver': 'auto',
'tol': 0.0,
'random_state': ''
}],
ransacparams={})
def taskGLM(subj,
task,
taskModel='canonical',
nuisModel='24pXaCompCorXVolterra',
nproc=8):
"""
This function runs a task-based GLM on a single subject
Will only regress out task-timing, using either a canonical HRF model or FIR model
Input parameters:
subj : subject number as a string
task : fMRI task name (2 runs per task for HCP data)
nuisModel : nuisance regression model (to identify input data)
nproc : number of processes to use via multiprocessing
"""
h5f = h5py.File(outputdir + subj + '_glmOutput_data.h5', 'a')
run1 = h5f[task + '_RL']['nuisanceReg_resid_' + nuisModel][:].copy()
run2 = h5f[task + '_LR']['nuisanceReg_resid_' + nuisModel][:].copy()
data = np.hstack((run1, run2))
h5f.close()
# Identify number of ROIs
nROIs = data.shape[0]
# Identify number of TRs
nTRs = data.shape[1]
# Load regressors for data
X = loadTaskTiming(subj, task, taskModel=taskModel, nRegsFIR=25)
taskRegs = X['taskRegressors'] # These include the two binary regressors
betas, resid = regression.regression(data.T, taskRegs, constant=True)
betas = betas.T # Exclude nuisance regressors
residual_ts = resid.T
h5f = h5py.File(outputdir + subj + '_glmOutput_data.h5', 'a')
outname1 = 'taskRegression/' + task + '_' + nuisModel + '_taskReg_resid_' + taskModel
outname2 = 'taskRegression/' + task + '_' + nuisModel + '_taskReg_betas_' + taskModel
try:
h5f.create_dataset(outname1, data=residual_ts)
h5f.create_dataset(outname2, data=betas)
except:
del h5f[outname1], h5f[outname2]
h5f.create_dataset(outname1, data=residual_ts)
h5f.create_dataset(outname2, data=betas)
h5f.close()
def test_LARS():
regress = regression(method=['LARS'],
yrange=[0.0, 100.0],
params=[{
'n_nonzero_coefs': 500,
'fit_intercept': True,
'positive': False,
'verbose': False,
'normalize': False,
'precompute': True,
'copy_X': True,
'eps': 2.220445,
'fit_path': True
}],
ransacparams={})
def test_Lasso_LARS():
regress = regression(method=['Lasso LARS'],
yrange=[0.0, 100.0],
params=[{
'alpha': 0.0,
'fit_intercept': True,
'positive': False,
'verbose': False,
'normalize': True,
'copy_X': True,
'precompute': 'Auto',
'max_iter': 500,
'eps': 2.220446,
'fit_path': True
}],
ransacparams={})
def test_Bayesian_Ridge():
regress = regression(method=['Bayesian Ridge'],
yrange=[0.0, 100.0],
params=[{
'n_iter': 300,
'tol': 0.001,
'alpha_1': 0.001,
'alpha_2': 1e-06,
'lambda_1': 1e-06,
'lambda_2': 1e-06,
'compute_score': False,
'fit_intercept': True,
'normalize': False,
'copy_X': True,
'verbose': False
}],
ransacparams={})
def test_SVR():
regress = regression(method=['SVR'],
yrange=[0.0, 100.0],
params=[{
'C': 1.0,
'epsilon': 0.1,
'kernel': 'rbf',
'degree': 0,
'gamma': 'auto',
'coef0': 0.0,
'shrinking': False,
'tol': 0.001,
'cache_size': 200,
'verbose': False,
'max_iter': -1
}],
ransacparams={})
def test_Elastic_Net():
regress = regression(method=['Elastic Net'],
yrange=[0.0, 100.0],
params=[{
'alpha': 1.0,
'l1_ratio': 0.5,
'fit_intercept': True,
'normalize': False,
'precompute': 'False',
'max_iter': 1000,
'copy_X': True,
'tol': 0.0001,
'warm_start': False,
'positive': False,
'selection': 'cyclic',
'random_state': 'None'
}],
ransacparams={})
def _nuisanceRegression(subj,
run,
inputdata,
outputdir,
model='24pXaCompCorXVolterra',
spikeReg=False,
zscore=False,
nproc=8):
"""
This function runs nuisance regression on the Glasser Parcels (360) on a single subjects run
Will only regress out noise parameters given the model choice (see below for model options)
Input parameters:
subj : subject number as a string
run : task run
outputdir: Directory for GLM output, as an h5 file (each run will be contained within each h5)
model : model choices for linear regression. Models include:
1. 24pXaCompCorXVolterra [default]
Variant from Ciric et al. 2017.
Includes (64 regressors total):
- Movement parameters (6 directions; x, y, z displacement, and 3 rotations) and their derivatives, and their quadratics (24 regressors)
- aCompCor (5 white matter and 5 ventricle components) and their derivatives, and their quadratics (40 regressors)
2. 18p (the legacy default)
Includes (18 regressors total):
- Movement parameters (6 directions) and their derivatives (12 regressors)
- Global signal and its derivative (2 regressors)
- White matter signal and its derivative (2 regressors)
- Ventricles signal and its derivative (2 regressors)
3. 16pNoGSR (the legacy default, without GSR)
Includes (16 regressors total):
- Movement parameters (6 directions) and their derivatives (12 regressors)
- White matter signal and its derivative (2 regressors)
- Ventricles signal and its derivative (2 regressors)
4. 12pXaCompCor (Typical motion regression, but using CompCor (noGSR))
Includes (32 regressors total):
- Movement parameters (6 directions) and their derivatives (12 regressors)
- aCompCor (5 white matter and 5 ventricle components) and their derivatives (no quadratics; 20 regressors)
5. 36p (State-of-the-art, according to Ciric et al. 2017)
Includes (36 regressors total - same as legacy, but with quadratics):
- Movement parameters (6 directions) and their derivatives and quadratics (24 regressors)
- Global signal and its derivative and both quadratics (4 regressors)
- White matter signal and its derivative and both quadratics (4 regressors)
- Ventricles signal and its derivative (4 regressors)
spikeReg : spike regression (Satterthwaite et al. 2013) [True/False]
Note, inclusion of this will add additional set of regressors, which is custom for each subject/run
zscore : Normalize data (across time) prior to fitting regression
nproc = number of processes to use via multiprocessing
"""
data = inputdata
tMask = np.ones((data.shape[1], ))
tMask[:framesToSkip] = 0
# Skip frames
data = data[:, framesToSkip:]
# Demean each run
data = signal.detrend(data, axis=1, type='constant')
# Detrend each run
data = signal.detrend(data, axis=1, type='linear')
tMask = np.asarray(tMask, dtype=bool)
nROIs = data.shape[0]
# Load nuisance regressors for this data
h5f = h5py.File(nuis_reg_dir + subj + '_nuisanceRegressors.h5', 'r')
if model == '24pXaCompCorXVolterra':
# Motion parameters + derivatives
motion_parameters = h5f[run]['motionParams'][:].copy()
motion_parameters_deriv = h5f[run]['motionParams_deriv'][:].copy()
# WM aCompCor + derivatives
aCompCor_WM = h5f[run]['aCompCor_WM'][:].copy()
aCompCor_WM_deriv = h5f[run]['aCompCor_WM_deriv'][:].copy()
# Ventricles aCompCor + derivatives
aCompCor_ventricles = h5f[run]['aCompCor_ventricles'][:].copy()
aCompCor_ventricles_deriv = h5f[run][
'aCompCor_ventricles_deriv'][:].copy()
# Create nuisance regressors design matrix
nuisanceRegressors = np.hstack(
(motion_parameters, motion_parameters_deriv, aCompCor_WM,
aCompCor_WM_deriv, aCompCor_ventricles,
aCompCor_ventricles_deriv))
quadraticRegressors = nuisanceRegressors**2
nuisanceRegressors = np.hstack(
(nuisanceRegressors, quadraticRegressors))
elif model == '18p':
# Motion parameters + derivatives
motion_parameters = h5f[run]['motionParams'][:].copy()
motion_parameters_deriv = h5f[run]['motionParams_deriv'][:].copy()
# Global signal + derivatives
global_signal = h5f[run]['global_signal'][:].copy()
global_signal_deriv = h5f[run]['global_signal_deriv'][:].copy()
# white matter signal + derivatives
wm_signal = h5f[run]['wm_signal'][:].copy()
wm_signal_deriv = h5f[run]['wm_signal_deriv'][:].copy()
# ventricle signal + derivatives
ventricle_signal = h5f[run]['ventricle_signal'][:].copy()
ventricle_signal_deriv = h5f[run]['ventricle_signal_deriv'][:].copy()
# Create nuisance regressors design matrix
tmp = np.vstack(
(global_signal, global_signal_deriv, wm_signal, wm_signal_deriv,
ventricle_signal, ventricle_signal_deriv
)).T # Need to vstack, since these are 1d arrays
nuisanceRegressors = np.hstack(
(motion_parameters, motion_parameters_deriv, tmp))
elif model == '16pNoGSR':
# Motion parameters + derivatives
motion_parameters = h5f[run]['motionParams'][:].copy()
motion_parameters_deriv = h5f[run]['motionParams_deriv'][:].copy()
# white matter signal + derivatives
wm_signal = h5f[run]['wm_signal'][:].copy()
wm_signal_deriv = h5f[run]['wm_signal_deriv'][:].copy()
# ventricle signal + derivatives
ventricle_signal = h5f[run]['ventricle_signal'][:].copy()
ventricle_signal_deriv = h5f[run]['ventricle_signal_deriv'][:].copy()
# Create nuisance regressors design matrix
tmp = np.vstack((wm_signal, wm_signal_deriv, ventricle_signal,
ventricle_signal_deriv
)).T # Need to vstack, since these are 1d arrays
nuisanceRegressors = np.hstack(
(motion_parameters, motion_parameters_deriv, tmp))
elif model == '12pXaCompCor':
# Motion parameters + derivatives
motion_parameters = h5f[run]['motionParams'][:].copy()
motion_parameters_deriv = h5f[run]['motionParams_deriv'][:].copy()
# WM aCompCor + derivatives
aCompCor_WM = h5f[run]['aCompCor_WM'][:].copy()
aCompCor_WM_deriv = h5f[run]['aCompCor_WM_deriv'][:].copy()
# Ventricles aCompCor + derivatives
aCompCor_ventricles = h5f[run]['aCompCor_ventricles'][:].copy()
aCompCor_ventricles_deriv = h5f[run][
'aCompCor_ventricles_deriv'][:].copy()
# Create nuisance regressors design matrix
nuisanceRegressors = np.hstack(
(motion_parameters, motion_parameters_deriv, aCompCor_WM,
aCompCor_WM_deriv, aCompCor_ventricles,
aCompCor_ventricles_deriv))
elif model == '36p':
# Motion parameters + derivatives
motion_parameters = h5f[run]['motionParams'][:].copy()
motion_parameters_deriv = h5f[run]['motionParams_deriv'][:].copy()
# Global signal + derivatives
global_signal = h5f[run]['global_signal'][:].copy()
global_signal_deriv = h5f[run]['global_signal_deriv'][:].copy()
# white matter signal + derivatives
wm_signal = h5f[run]['wm_signal'][:].copy()
wm_signal_deriv = h5f[run]['wm_signal_deriv'][:].copy()
# ventricle signal + derivatives
ventricle_signal = h5f[run]['ventricle_signal'][:].copy()
ventricle_signal_deriv = h5f[run]['ventricle_signal_deriv'][:].copy()
# Create nuisance regressors design matrix
tmp = np.vstack(
(global_signal, global_signal_deriv, wm_signal, wm_signal_deriv,
ventricle_signal, ventricle_signal_deriv
)).T # Need to vstack, since these are 1d arrays
nuisanceRegressors = np.hstack(
(motion_parameters, motion_parameters_deriv, tmp))
quadraticRegressors = nuisanceRegressors**2
nuisanceRegressors = np.hstack(
(nuisanceRegressors, quadraticRegressors))
if spikeReg:
# Obtain motion spikes
try:
motion_spikes = h5f[run]['motionSpikes'][:].copy()
nuisanceRegressors = np.hstack((nuisanceRegressors, motion_spikes))
except:
print 'Spike regression option was chosen... but no motion spikes for subj', subj, '| run', run, '!'
# Update the model name - to keep track of different model types for output naming
model = model + '_spikeReg'
if zscore:
model = model + '_zscore'
h5f.close()
# Skip first 5 frames of nuisanceRegressors, too
nuisanceRegressors = nuisanceRegressors[framesToSkip:, :].copy()
betas, resid = regression.regression(data.T,
nuisanceRegressors,
constant=True)
betas = betas.T # Exclude nuisance regressors
residual_ts = resid.T
if zscore:
residual_ts = stats.zscore(residual_ts, axis=1)
outname1 = run + '/nuisanceReg_resid_' + model
outname2 = run + '/nuisanceReg_betas_' + model
outputfilename = outputdir + subj + '_glmOutput_data.h5'
h5f = h5py.File(outputfilename, 'a')
try:
h5f.create_dataset(outname1, data=residual_ts)
h5f.create_dataset(outname2, data=betas)
except:
del h5f[outname1], h5f[outname2]
h5f.create_dataset(outname1, data=residual_ts)
h5f.create_dataset(outname2, data=betas)
h5f.close()
def taskGLM_onRest(subj,
taskModel='canonical',
nuisModel='24pXaCompCorXVolterra',
nproc=8):
"""
This function runs a task-based GLM on a single subject
Will only regress out task-timing, using either a canonical HRF model or FIR model
Input parameters:
subj : subject number as a string
task : fMRI task name (2 runs per task for HCP data)
nuisModel : nuisance regression model (to identify input data)
nproc : number of processes to use via multiprocessing
"""
h5f = h5py.File(outputdir + subj + '_glmOutput_data.h5', 'a')
restname = 'rfMRI_REST'
run1 = h5f[restname + '1_RL']['nuisanceReg_resid_' + nuisModel][:].copy()
run2 = h5f[restname + '1_LR']['nuisanceReg_resid_' + nuisModel][:].copy()
run3 = h5f[restname + '2_RL']['nuisanceReg_resid_' + nuisModel][:].copy()
run4 = h5f[restname + '2_LR']['nuisanceReg_resid_' + nuisModel][:].copy()
data = np.hstack((run1, run2, run3, run4))
h5f.close()
# Identify number of ROIs
nROIs = data.shape[0]
# Identify number of TRs
nTRs = data.shape[1]
# Load regressors for data, for each task, and
trcount = 0
data_resids = []
for task in taskNames:
X = loadTaskTiming(subj, task, taskModel=taskModel, nRegsFIR=25)
taskRegs = X['taskRegressors']
trstart = trcount
trend = trstart + taskRegs.shape[0]
betas, resid = regression.regression(data[:, trstart:trend].T,
taskRegs,
constant=True)
betas = betas.T # Exclude nuisance regressors
residual_ts = resid.T
data_resids.extend(residual_ts.T)
trcount = trend
# Append the rest of the resting-state time series, just in case
# trstart = trcount
# data_resids.extend(data[:,trstart:].T)
data_resids = np.asarray(data_resids)
residual_ts = data_resids.T
h5f = h5py.File(outputdir + subj + '_glmOutput_data.h5', 'a')
outname1 = 'taskRegression/' + restname + '_' + nuisModel + '_taskReg_resid_' + taskModel
try:
h5f.create_dataset(outname1, data=residual_ts)
except:
del h5f[outname1]
h5f.create_dataset(outname1, data=residual_ts)
h5f.close()
return residual_ts
