# Built-in imports
import math
import ctypes
# External imports
import numpy as np
from scipy.spatial.distance import pdist
from scipy.signal import decimate
# Medusa imports
from medusa.components import ThreadWithReturnValue
from medusa.utils import check_dimensions
[docs]def central_tendency_measure(signal, r):
"""
This method implements the central tendency measure (CTM). This parameter is
useful to quantify the variability of a signal. It is based on calculating
the second-order differences diagram of the time series and then counting
the points within a radius "r". CTM assigns higher values to less variable
signals
References
----------
Cohen, M. E., Hudson, D. L., & Deedwania, P. C. (1996). Applying
continuous chaotic modeling to cardiac signal analysis. IEEE Engineering in
Medicine and Biology Magazine, 15(5), 97-102.
Parameters
----------
signal : numpy.ndarray
Signal with shape [n_epochs, n_samples, n_channels].
r : double
Radius used to compute the CTM.
Returns
-------
ctm : numpy.ndarray
CTM values for each channel in "signal". [n_epochs, n_channels].
"""
# Error check
if not np.issubdtype(signal.dtype, np.number):
raise ValueError('data matrix contains non-numeric values')
# Check dimensions
signal = check_dimensions(signal)
# Signal dimensions
n_epo = signal.shape[0]
n_samp = signal.shape[1]
n_cha = signal.shape[2]
# Values within a range (mean +- 3 std)
upper_bound = np.mean(signal, axis=1) + 3 * np.std(signal, axis=1)
lower_bound = np.mean(signal, axis=1) - 3 * np.std(signal, axis=1)
idx_within_range = np.logical_and((signal < upper_bound[:, None, :]),
(signal > lower_bound[:, None, :]))
idx_out_upper = (signal > upper_bound[:, None, :])
idx_out_lower = (signal < lower_bound[:, None, :])
# Maximum value in the above defined range
max_value = np.empty((n_epo, n_cha))
for ep in range(n_epo):
for ch in range(n_cha):
max_value[ep, ch] = np.max(
abs(signal[ep, idx_within_range[ep, :, ch], ch]), axis=0)
# Normalize the values within the range by its maximum.Values above that
# range will be 1, and below the range will be - 1
data_norm = np.zeros_like(signal)
data_norm[idx_within_range] = np.divide(
signal[idx_within_range],
np.tile(max_value, (1, n_samp, 1)).flatten()[
idx_within_range.flatten()])
data_norm[idx_out_upper] = 1
data_norm[idx_out_lower] = -1
# Difference time series
y = data_norm[:, 3:n_samp, :] - data_norm[:, 2:n_samp - 1, :]
x = data_norm[:, 2:n_samp - 1, :] - data_norm[:, 1:n_samp - 2, :]
# CTM - Values below the radius 'r'
ctm = np.sum(np.sqrt(np.square(x) + np.square(y)) < r, axis=1) / (
n_samp - 2)
return ctm
[docs]def sample_entropy(signal, m, r, dist_type='chebyshev'):
""" This method implements the sample entropy (SampEn). SampEn is an
irregularity measure that assigns higher values to more irregular time
sequences. It has two tuning parameters: the sequence length (m) and the
tolerance (r)
Notes: IF A = 0 or B = 0, SamEn would return an infinite value.
However, the lowest non-zero conditional probability that SampEn should
report is A/B = 2/[(N-m-1)*(N-m)].
SampEn has the following limits:
- Lower bound: 0
- Upper bound : log(N-m) + log(N-m-1) - log(2)
References
----------
Richman, J. S., & Moorman, J. R. (2000). Physiological
time-series analysis using approximate entropy and sample entropy. American
Journal of Physiology-Heart and Circulatory Physiology.
Parameters
----------
signal : numpy.ndarray
Signal with shape [n_epochs, n_samples, n_channels].
m : int
Sequence length
r : float
Tolerance
dist_type : string
Distance metric
Returns
-------
sampen : numpy.ndarray
SampEn value. [n_epochs, n_channels]
"""
# Check dimensions
signal = check_dimensions(signal)
# Check Errors
if m > signal.shape[1]:
raise ValueError('Embedding dimension must be smaller than the signal '
'length (m<N).')
if not isinstance(dist_type, str):
raise ValueError('Distance type must be a string.')
if dist_type not in ['braycurtis', 'canberra', 'chebyshev', 'cityblock',
'correlation', 'cosine', 'dice', 'euclidean',
'hamming', 'jaccard', 'jensenshannon', 'kulsinski',
'mahalanobis', 'matching', 'minkowski',
'rogerstanimoto', 'russellrao', 'seuclidean',
'sokalmichener', 'sokalsneath', 'sqeuclidean', 'yule']:
raise ValueError(
'Distance type unknown. Please, check allowed distances'
'in pdist function from scipy.spatial.distance module.')
# Useful parameters
n_epo = signal.shape[0]
N = signal.shape[1]
n_channels = signal.shape[2]
sigma = np.std(signal, axis=1)
templates_m = []
templates_m_plus_one = []
B, A, value = np.empty((n_epo, n_channels)), np.empty((n_epo, n_channels)),\
np.empty((n_epo, n_channels))
# Calculate B values
for i in range(N - m + 1):
templates_m.append(signal[:, i:i + m, :])
templates_m = np.array(templates_m)
for e_idx in range(n_epo):
w_threads = []
for ch_idx in range(n_channels):
t = ThreadWithReturnValue(
target=pdist,
args=(templates_m[:, e_idx, :, ch_idx], dist_type,))
w_threads.append(t)
t.start()
for th_idx, thread in enumerate(w_threads):
B[e_idx, th_idx] = np.sum(thread.join() <= sigma[0, th_idx] * r)
# Check if there is any B = 0
zeros_idx = np.where(B == 0)
value[zeros_idx] = math.inf
# Calculate A values
m += 1
for i in range(N - m + 1):
templates_m_plus_one.append(signal[:, i:i + m, :])
templates_m_plus_one = np.array(templates_m_plus_one)
for e_idx in range(n_epo):
w_threads = []
for ch_idx in range(n_channels):
t = ThreadWithReturnValue(
target=pdist, args=(templates_m_plus_one[:, e_idx, :, ch_idx],
dist_type,))
w_threads.append(t)
t.start()
for th_idx, thread in enumerate(w_threads):
A[e_idx, th_idx] = np.sum(thread.join() <= sigma[0, th_idx] * r)
# Check if there is any A = 0
zeros_idx = np.where(A == 0)
value[zeros_idx] = math.inf
non_inf = np.where(value != math.inf)
value[non_inf] = -np.log((A[non_inf] / B[non_inf])*((N - m + 1) / (N - m - 1)))
# If there is infinity values
inf_indx = np.where(value == math.inf)
value[inf_indx] = -np.log(2 / ((N - m - 1) * (N - m)))
return value
[docs]def multiscale_entropy(signal, max_scale, m, r):
""" This method implements the Multiscale Entropy (MSE). MSE is a
measurement of complexity which measures the irregularity of a signal over
multiple time-scales. This is accomplished through estimation of the Sample
Entropy (SampEn) on coarse-grained versions of the original signal. As a
result of the these calculations, MSE curves are obtained and can be used to
compare the complexity of time-series. The MSE curve whose entropy values
are higher for the most of time-scales is considered more complex.
References
----------
Costa, M., Goldberger, A. L., & Peng, C. K. (2005). Multiscale entropy
analysis of biological signals, Physical review E, 71(2), 021906.
Parameters
----------
signal : numpy.ndarray
MEEG Signal. [n_epochs, n_samples, n_channels].
max_scale : int
Maximum scale value
m : int
Sequence length
r : float
Tolerance
Returns
-------
mse_result : numpy.ndarray
SampEn values for each scale for each channel in "signal" .
[n_epochs, max_scale, n_channels].
"""
# Check dimensions
signal = check_dimensions(signal)
# Signal dimensions
n_epo = signal.shape[0]
n_channels = signal.shape[2]
mse_result = np.empty((n_epo, max_scale, n_channels))
w_threads = list()
scales = list()
for i in range(1, max_scale + 1):
if i == 1:
t = ThreadWithReturnValue(
target=sample_entropy,
args=(signal, m, r, 'chebyshev'))
else:
t = ThreadWithReturnValue(
target=sample_entropy,
args=(__coarse_grain(signal, i),
m, r, 'chebyshev'))
w_threads.append(t)
scales.append(i)
t.start()
for t_idx, thread in enumerate(w_threads):
mse_result[:, t_idx, :] = thread.join()
return mse_result
def __coarse_grain(signal, scale, decimate_mode=True):
if decimate_mode:
return decimate(signal, scale, axis=1)
else:
# Signal dimensions
n_epo = signal.shape[0]
N = signal.shape[1]
n_cha = signal.shape[2]
# Number of coarse grains in which the signal is split
tau = int(round(N / scale))
# Returned signal
y = np.empty((n_epo,tau,n_cha))
for i in range(tau):
y[:, i, :] = np.mean(signal[:, i * scale:(i * scale + scale), :])
return y
def __lempelziv_complexity(signal):
""" Calculate the signal binarisation and the Lempel-Ziv's complexity. This
version allows the algorithm to be used for signals with more than one
channel at the same time.
Warnings: This function is deprecated and will be removed in future updates
of MEDUSA Kernel.
Parameters
---------
signal: numpy 2D array
Signal with shape [n_samples, n_channels]
Returns
-------
lz_channel_values: numpy 1D matrix
Lempel-Ziv values for each channel with shape [n_channels].
"""
if signal.size == len(signal):
signal = signal[:, np.newaxis]
signal = __binarisation(signal, [signal.shape[0]], signal.shape[0])
if signal.shape[1] == 1:
return __lz_algorithm(signal)
else:
lz_channel_values = np.empty((signal.shape[1]))
working_threads = list()
for ch in range(signal.shape[1]):
t = ThreadWithReturnValue(target=__lz_algorithm,
args=(signal[:, ch],))
working_threads.append(t)
t.start()
for index, thread in enumerate(working_threads):
lz_channel_values[index] = thread.join()
return lz_channel_values
[docs]def lempelziv_complexity(signal):
"""Calculate the signal binarisation and the Lempel-Ziv's complexity.
This function takes advantage of its implementation in C to achieve a high
performance.
Parameters
---------
signal: numpy.ndarray
MEEG Signal [n_epochs, n_samples, n_channels]
Returns
-------
lz_result: numpy.ndarray
Lempel-Ziv values for each epoch and each channel. [n_epochs, n_channels].
"""
# Check dimensions
signal = check_dimensions(signal)
# Signal dimensions
n_epo = signal.shape[0]
n_sample = signal.shape[1]
n_cha = signal.shape[2]
# Initialize result matrix
lz_result = np.empty((n_epo, n_cha))
# Adapt the signal to the format required by the DLL
signal = np.reshape(np.transpose(signal, (0, 2, 1)), (n_epo, -1))
# Get function from dll
dll_file = ".\computeLZC.dll"
lib = ctypes.cdll.LoadLibrary(dll_file)
lzc_func = lib.computeLempelZivCmplx # Access function
for e_idx, epochs in enumerate(signal):
# Create empty output vector
lz_channel_values = np.zeros(int(n_cha), dtype=np.double)
# Define inputs and outputs
lzc_func.restype = None
array_1d_double = np.ctypeslib.ndpointer(dtype=np.double, ndim=1,
flags='CONTIGUOUS')
lzc_func.argtypes = [array_1d_double, array_1d_double, ctypes.c_int32,
ctypes.c_int32]
# Call the function in the dll
lzc_func(signal[e_idx, :], lz_channel_values, int(n_sample * n_cha),
int(n_sample))
lz_result[e_idx, :] = lz_channel_values
return lz_result
[docs]def multiscale_lempelziv_complexity(signal, W):
"""Calculate the multiscale signal binarisation and the Multiscale
Lempel-Ziv's complexity.
References
----------
Ibáñez-Molina, A. J., Iglesias-Parro, S., Soriano, M. F., & Aznarte, J. I,
Multiscale Lempel-Ziv complexity for EEG measures. Clinical Neurophysiology,
(2015), 126(3), 541–548.
Parameters
---------
signal: numpy.ndarray
MEEG Signal [n_epochs, n_samples, n_channels]
W: list or numpy 1D array
Set of window length to consider at multiscale binarisation stage.
Values must be odd.
Returns
-------
result: numpy.ndarray
Matrix of results with shape [n_epochs, n_window_length, n_channels]
"""
# Check dimensions
signal = check_dimensions(signal)
# Signal dimensions
n_epo = signal.shape[0]
n_cha = signal.shape[2]
# Useful parameter
w_max = W[-1] + (W[1] - W[0])
# Define a matrix to store results
result = np.empty((n_epo, len(W), n_cha))
# First get binarised signal
for ep_idx, epoch in enumerate(signal):
for w_idx, w in enumerate(W):
binarised_signal = __binarisation(epoch, w, w_max, multiscale=True)
# Parallelize the calculations if n_channel > 1
if binarised_signal.shape[1] > 1:
threads = []
for ch in range(binarised_signal.shape[1]):
t = ThreadWithReturnValue(target=__lz_algorithm,
args=(binarised_signal[:, ch],))
threads.append(t)
t.start()
for ch_idx, t in enumerate(threads):
result[ep_idx, w_idx, ch_idx] = t.join()
else:
result[ep_idx, w_idx, :] = __lz_algorithm(binarised_signal)
return result
def __lz_algorithm(signal):
"""
Lempel-Ziv's complexity algorithm implemented in Python.
References
----------
F. Kaspar, H. G. Schuster, "Easily-calculable measure for the complexity of
spatiotemporal patterns", Physical Review A, Volume 36, Number 2 (1987).
Parameters
---------
signal: numpy 1D array
Signal with shape of [n_samples]
Returns
-------
value: float
Result of algorithm calculations.
"""
signal = signal.flatten().tolist()
i, k, l = 0, 1, 1
c, k_max = 1, 1
n = len(signal)
while True:
if signal[i + k - 1] == signal[l + k - 1]:
k = k + 1
if l + k > n:
c = c + 1
break
else:
if k > k_max:
k_max = k
i = i + 1
if i == l:
c = c + 1
l = l + k_max
if l + 1 > n:
break
else:
i = 0
k = 1
k_max = 1
else:
k = 1
value = c * (np.log2(n) / n)
return value
def __binarisation(signal, w_length, w_max, multiscale=False):
"""
This function returns a binarised version of original signal. It can be used
in both multiscale and simple binarisations. If multiscale mode is chosen,
signals are shortened so that they all have the same length, taking the
maximum window as a reference. Binarisation is performed by means of a
median-based comparison.
Parameters
----------
signal: numpy 2D array
Signal with shape [n_samples x n_channels]
w_length: int
Window length to perform multiscale binarisation
w_max: int
Value of the maximum window length
multiscale: bool
If is True, performs the multiscale binarisation
Returns
-------
signal_binarised: numpy.ndarray
Signal binarised with shape of [n_samples_shortened x n_channels]
The n_samples_shortened parameter is calculated from w_max to ensure that
all signals have the same length, regardless of the window length used.
"""
if multiscale:
if w_length is None:
raise ValueError('Width of window must be an integer value')
if w_length % 2 == 0:
raise ValueError('Width of window must be an odd value.')
if w_max is None:
raise ValueError('Maximum window width must be an integer value')
# Get smoothed version from original signal
smoothed = __multiscale_median_threshold(signal, w_length)
# Useful parameters
half_wind = int((w_length - 1) / 2)
max_length = signal.shape[0] + 1 - w_max
length_diff = smoothed.shape[0] - max_length
# Shorten original and smoothed version
smoothed_shortened = \
smoothed[int(length_diff / 2):-int(length_diff / 2), :]
signal_shortened = \
signal[half_wind: signal.shape[0] - half_wind, :]
signal_shortened = \
signal_shortened[int(length_diff / 2):-int(length_diff / 2), :]
# Define template of binarised signal
signal_binarised = \
np.zeros((signal_shortened.shape[0], signal_shortened.shape[1]))
# Binarise the signal
idx_one = signal_shortened >= smoothed_shortened
signal_binarised[idx_one] = 1
else:
signal_binarised = np.zeros((len(signal), signal.shape[1]))
median = np.median(signal, axis=0)
idx_one = signal >= median
signal_binarised[np.squeeze(idx_one)] = 1
return signal_binarised
def __multiscale_median_threshold(signal, w_length):
""" Signal smoothing function. For each sample, we define a window in which
the sample is in centre position. The median value of the window is
calculated and assigned to this sample to obtain a smoothed version of the
original signal.
References
----------
Ibáñez-Molina, A. J., Iglesias-Parro, S., Soriano, M. F., & Aznarte, J. I,
Multiscale Lempel-Ziv complexity for EEG measures. Clinical Neurophysiology,
(2015), 126(3), 541–548.
Parameters
---------
signal: numpy 2D array
Signal with shape of [n_samples, n_channels]
w_length: int
Length of window to calculate median values in smoothing process
Returns
-------
smoothed_signal: numpy 2D array
Smoothed version with shape of [n_samples + 1 - w_length, n_channels]
As a result of windowing, the final samples of the signal are lost.
"""
# Template of smoothed signal
smoothed_signal = np.zeros((
signal.shape[0] + 1 - w_length, signal.shape[1]))
half_wind = int((w_length - 1) / 2)
# Index of sample to be smoothed from median window value
index = 0
# We define a window with samp in central position and
# get median value to smooth original signal
for samp in range(half_wind, signal.shape[0] - half_wind):
smoothed_signal[index, :] = np.median(
signal[samp - half_wind: samp + half_wind], axis=0)
index += 1
return smoothed_signal