def test_electrodes_selection(self):
dd = cp.Data(np.random.randn(5, 100, 4), fs=10)
with self.assertRaises(ValueError):
dd.select_channels([0, 2, 5])
dd.select_channels([0, 1, 3])
self.assertEquals(dd.channelsnr, 3)
self.assertEquals(dd.data.shape[0], 3)
def significance(self,
max_order,
method="DTF",
order=None,
signf_threshold=0.05,
Nrep=200,
alpha=0.05):
"""Compute and plot the binary matrix having as positive elements
the related p-values of the significance matrix less than threshold.
Args:
-----------
method : string
"PDC" or "EDF". Estimation method for connectivity.
order : int
Value of order of autoregressive multivariate model.
max_order : int
Max order computable by akaike algorithm.
signf_threshold : float
Threshold for p-values of the significance matrix.
Nrep : int
Number of resamples for the computation of the significance matrix.
alpha : float
Type 1 error rate for the significance level.
"""
if not order:
best, crit = cp.Mvar.order_akaike(self.values, max_order)
plt.plot(1 + np.arange(len(crit)),
crit,
marker='o',
linestyle='dashed',
markersize=8,
markerfacecolor='yellow')
plt.grid()
plt.show()
p = best
else:
p = order
print('Best model order p: {}'.format(p))
data = cp.Data(self.values, chan_names=self.channels)
data.fit_mvar(p, 'yw')
if method == 'DTF':
matrix_values = data.conn('dtf')
else:
matrix_values = data.conn('pdc')
significance_matrix = data.significance(Nrep=Nrep,
alpha=alpha,
verbose=False)
significance_matrix[significance_matrix < 0.05] = 1
significance_matrix[significance_matrix != 1] = 0
self.significance_matrix = significance_matrix
plt.imshow(significance_matrix, cmap='Greys', interpolation='nearest')
plt.show()
def fit_model(data, fs, resolution, method, freq=None, boot=False):
'''
Fit an MVAR model, and compute connecitvity estimation via PDC or DTF.
'''
if boot:
data = connectivipy.Data(data=data,
fs=160,
chan_names=get_labels_nodes(19))
data.fit_mvar(method="yw")
data.conn("pdc")
res = data.significance(Nrep=200, alpha=0.05, verbose=False)
np.fill_diagonal(res, 0)
return res
model = connectivipy.Mvar().fit(data, method="yw")
if method == "dtf":
if freq:
res = connectivipy.conn.dtf_fun(model[0],
model[1],
fs=fs,
resolution=resolution)[freq, :, :]
np.fill_diagonal(res, 0)
return res
else:
return connectivipy.conn.dtf_fun(model[0],
model[1],
fs=fs,
resolution=resolution)
elif method == "pdc":
if freq:
res = connectivipy.conn.pdc_fun(model[0],
model[1],
fs=fs,
resolution=resolution)[freq, :, :]
np.fill_diagonal(res, 0)
return res
else:
return connectivipy.conn.pdc_fun(model[0],
model[1],
fs=fs,
resolution=resolution)
else:
return "Wrong method. Use \"pdc\" or \"dtf\""
def s(self):
table = pd.read_csv(
self.txt_name,
delimiter='\s+',
)
labels = list(table.label)
for i in range(len(labels)):
# clean labels name
labels[i] = labels[i].replace('..', '')
labels[i] = labels[i].replace('.', '')
sub_data, idx_19 = self.sub_data()
labels_19 = [labels[i] for i in idx_19]
data = cp.Data(sub_data, fs=self.fs, chan_names=labels)
data.fit_mvar()
pdc_values = data.conn('pdc')
s_matrix = data.significance(Nrep=200, alpha=0.05)
return s_matrix
def significance(self,signf_threshold=None,channels=[],order=None,Nrep=200,alpha=0.05,visual=False,path=None,freq=10,name='significance'):
"""Computes the significance (p-value) of the assoiated nodes connections using resampling method for the graph created
Inputs:
- signf_threshold: limits of significance under which we want to exclude specific nodes
- channels: list of the channels we want to extract for the rest of the procedure, if empty we use them all.
- order: order of the MVAR model
- Nrep: number od resamples - number of repetitions for the resampling algorithm
- alpha: (default 0.05) - type I error rate (significance level)
- visual: if TRUE plots the new network
- path: path of .edf file we want to import the data if visual TRUE
- freq: sample frequency if visual TRUE
- name: name of the output plot of the network if visual TRUE
"""
df = self.df
if channels:
df = df[channels]
self.values = df.T.values
self.channels = df.columns.values
self.num_of_channels,self.num_of_samples = self.values.shape
data = cp.Data(self.values,chan_names=self.channels)
if order:
self.p = order
data.fit_mvar(self.p,self.conn_algorithm)
if self.method == 'DTF':
matrix_values = data.conn('dtf')
else:
matrix_values = data.conn('pdc')
self.significance_matrix = data.significance(Nrep=Nrep, alpha=alpha,verbose=False)
if signf_threshold:
elim_indices = np.argwhere(self.significance_matrix>signf_threshold)
try:
self.import_data(path,channels = df.columns.values)
self.connectivity(freq=freq,significance=elim_indices,order=self.p,threshold=round(self.density,1))
if visual:
#self.import_data(path,channels = df.columns.values)
#self.connectivity(freq=freq,significance=elim_indices,order=self.p,threshold=round(self.density,1))
self.show_graph(name)
except:
print("Oops! That was not a valid significance threshold number. Try again...")
def compute_connectivity(self,
freq,
method="PDC",
order=None,
max_order=10,
plot=False,
resolution=100,
threshold=None):
"""Compute the connectivity matrix and the binary matrix of the EEG data,
using PDC or DTF method for the estimation.
Args:
-----------
freq : int
Frequency value for the connectivity matrix.
method : string
"PDC" or "EDF". Estimation method for connectivity.
order : int
Value of order of autoregressive multivariate model.
max_order : int
Max order computable by akaike algorithm.
plot : boolean
Whether to plot the akaike algo results or not.
resolution : int
Number of spectrum datapoints.
threshold : float
Density threshold for the computation of the connectivity matrix.
Between 0 and 1.
"""
if not order:
best, crit = cp.Mvar.order_akaike(self.values, max_order)
p = best
if plot:
plt.plot(1 + np.arange(len(crit)),
crit,
marker='o',
linestyle='dashed',
markersize=8,
markerfacecolor='yellow')
plt.grid()
plt.show()
print('Best model order p: {}'.format(best))
else:
p = order
data = cp.Data(self.values, chan_names=self.channels)
data.fit_mvar(p, "yw")
# multivariate model coefficient (see slides)
ar, vr = data.mvar_coefficients
if method == 'DTF':
Adj = cp.conn.dtf_fun(ar, vr, fs=self.sample_freq,
resolution=100)[freq, :, :]
else:
Adj = cp.conn.pdc_fun(ar, vr, fs=self.sample_freq,
resolution=100)[freq, :, :]
np.fill_diagonal(Adj, 0)
# create Graph from Adj matrix
G = nx.from_numpy_matrix(np.array(Adj), create_using=nx.DiGraph)
A = nx.adjacency_matrix(G)
A = A.toarray()
# set values of diagonal zero to avoid self-loops
np.fill_diagonal(A, 0)
# reduce Graph density
while (nx.density(G) > threshold):
# find min values different from zeros
arg_min = np.argwhere(A == np.min(A[np.nonzero(A)]))
i, j = arg_min[0][0], arg_min[0][1]
# remove i,j edge from the graph
G.remove_edge(i, j)
# recalculate the graph
A = nx.adjacency_matrix(G)
A = A.toarray()
np.fill_diagonal(A, 0)
# np.fill_diagonal(A,diag)
density = nx.density(G)
connectivity_matrix = A.copy()
A[A > 0] = 1
binary_adjacency_matrix = A
self.connectivity_matrix = connectivity_matrix
self.binary_adjacency_matrix = binary_adjacency_matrix
# load coordinates
self.load_channel_locations()
# create directed binary graph
G = nx.DiGraph(binary_adjacency_matrix)
new_labels = {}
for i, node in enumerate(G.nodes):
new_labels[node] = self.channels[i]
self.G = nx.relabel.relabel_nodes(G, new_labels, copy=True)
# nx.set_node_attributes(self.G, self.channel_locations, "pos")
# create directed weighted graph
Gw = nx.DiGraph(connectivity_matrix)
self.Gw = nx.relabel.relabel_nodes(Gw, new_labels, copy=True)
def connectivity(self,freq,algorithm='yw',order=None,max_order=10,plot=False,resolution=100,threshold=None,mode=0,significance=[]):
"""
Computes the connectivity matrix of a graph using a specific connectivity method (DTF or PDC) and MVar model fitting
Inputs:
-freq: sample frequency
-algorithm: default Yule-Walker algorithm
-order: MVAR model order
-max_order: Maximum order to compute the best model's order
-plot: (DEFAULT: FALSE) if TRUE, plotting of the model order respect to minimization of BIC critirion
-resolution: frequency resolution
-threshold: (float between 0,1) percentage of density threshold
-mode: (default: 0 for Directed graph, else 1 for Undirected graph)
-significance: list with the nodes we want to exclude - filter out from our analysis regarding their significance value
"""
self.conn_algorithm = algorithm
if not order:
#best,crit = cp.Mvar.order_schwartz(self.values,max_order) ##BIC
best,crit = cp.Mvar.order_akaike(self.values,max_order) ##AIC
if plot:
plt.plot(1+np.arange(len(crit)), crit,marker='o', linestyle='dashed',markersize=8,markerfacecolor='yellow')
plt.grid()
plt.show()
self.p = best
print()
print('best model order p: {}'.format(best))
print()
else:
self.p = order
data = cp.Data(self.values,chan_names=self.channels)
data.fit_mvar(self.p, self.conn_algorithm)
ar, vr = data.mvar_coefficients
if self.method == 'DTF':
conn_matrix = cp.conn.DTF()
else:
conn_matrix = cp.conn.PDC()
Adj = conn_matrix.calculate(ar,vr,self.freq_sample,resolution)
self.res = np.linspace(0,self.freq_sample/2,resolution) ## matrix with resolution of frequencies
## choosing frequency equal to the sample_frequency
Adj = Adj[np.where(self.res == self.find_closest_freq(freq)),:,:].reshape(self.num_of_channels,self.num_of_channels)
###############################################################################
np.fill_diagonal(Adj,0)
if len(significance) > 0:
for a,b in significance:
Adj[a,b] = 0
################################################################################
self.G = Graph.Weighted_Adjacency(Adj.tolist(), mode = mode)
self.G.vs["label"] = list(map(lambda x: re.sub('\.', '', x), self.channels))
locations = pd.read_csv("./data/channel_locations.csv")
coords = {k[0]: (k[1],k[2]) for k in locations.values}
self.G.vs["coords"] = [coords[k["label"]] for k in self.G.vs]
A = np.array(self.G.get_adjacency(attribute = "weight").data)
diag = np.diag(A)
# set values of diagonal zero to avoid self-loops
np.fill_diagonal(A,0)
if threshold:
while(self.G.density() > threshold):
arg_min = np.argwhere(A == np.min(A[np.nonzero(A)]))
i,j = arg_min[0][0],arg_min[0][1]
self.G.delete_edges([(i,j)])
A = np.array(self.G.get_adjacency(attribute = "weight").data)
np.fill_diagonal(A,0)
np.fill_diagonal(A,diag)
self.density = self.G.density()
self.connectivity_matrix = A.copy()
A[A>0] = 1
self.binary_adjacency_matrix = A
self.G.vs['degree'] = self.G.degree()
A[0, 3, 4] = 0.25 * 2**0.5
A[0, 4, 3] = -0.25 * 2**0.5
A[0, 4, 4] = 0.25 * 2**0.5
# multitrial signal generation from a matrix above
# let's generate 5-channel signal with 1000 data points
# and 5 trials using function mvar_gen
ysig = np.zeros((5, 10**3, 5))
ysig[:, :, 0] = mvar_gen(A, 10**3)
ysig[:, :, 1] = mvar_gen(A, 10**3)
ysig[:, :, 2] = mvar_gen(A, 10**3)
ysig[:, :, 3] = mvar_gen(A, 10**3)
ysig[:, :, 4] = mvar_gen(A, 10**3)
#### connectivity analysis
data = cp.Data(ysig, 128, ["Fp1", "Fp2", "Cz", "O1", "O2"])
# you may want to plot data (in multitrial case only one trial is shown)
data.plot_data()
# fit mvar using Yule-Walker algorithm and order 2
data.fit_mvar(2, 'yw')
# you can capture fitted parameters and residual matrix
ar, vr = data.mvar_coefficients
# now we investigate connectivity using gDTF
gdtf_values = data.conn('gdtf')
gdtf_significance = data.significance(Nrep=200, alpha=0.05)
data.plot_conn('gDTF')
def test_mvar_calc(self):
data = cp.Data(ys, 128, ["Fp1", "Fp2", "Cz", "O1", "O2"])
data.fit_mvar(2, 'vm')
acoef, vcoef = data.mvar_coefficients
self.assertEquals(acoef.shape, (2, 5, 5))
self.assertEquals(vcoef.shape, (5, 5))
def test_conn2(self):
dat = cp.Data(ys)
dat.fit_mvar(2, 'vm')
estm = dat.short_time_conn('dtf', nfft=100, no=10)
stst = dat.short_time_significance(Nrep=100, alpha=0.5, verbose=False)
self.assertTrue(np.all(stst <= 1))
def test_conn(self):
dat = cp.Data(ys)
with self.assertRaises(AttributeError):
dat.conn('dtf')
def test_resample(self):
dd = cp.Data(np.random.randn(3, 100, 4), fs=10)
dd.resample(5)
self.assertEquals(dd.fs, 5)