def smooth(self, data):
sm_size = int(0.08 * self.sample_rate)
smoothed_data, _ = st.smoother(signal=data,
kernel='hamming',
size=sm_size, mirror=True)
return smoothed_data
def diffSCR(data,fs,min_amplitude): #onset at local minimum; peak at diff maximum df = np.diff(data) df = np.append(df,df[-1]) size = int(1. * fs) df = tools.smoother(df,'bartlett',size)['signal'] #plotsignal([data,df],False) onsets, pks, amps = find_zeropeak(df,min_amplitude,'diff') return onsets,pks,amps,df
def compute_intervals(peaks, smooth=False, size=3):
intervals = [0]
for i in range(len(peaks)-1):
intervals.append(peaks[i+1]-peaks[i])
intervals = np.array(intervals)
if smooth and (len(intervals) > 1):
intervals, _ = smoother(signal=intervals, kernel='boxcar', size=size, mirror=True)
return intervals
def read_sample(path, name, sampling_rate, preprocess=True):
dataset = pandas.read_csv(path + name + '.txt',
delimiter='\t',
skiprows=4,
skipfooter=1,
engine='python')
x_signal = dataset.values[:total_length * sampling_rate, 0]
y = dataset.values[:total_length * sampling_rate, 1]
y_signal = normalize_data(pandas.DataFrame(y), max_range, min_range)
if preprocess is True:
smoothed_signal, params = smoother(y_signal[:, 0])
y_signal = np.reshape(smoothed_signal, [smoothed_signal.shape[0], 1])
if sampling_rate != rate:
down_sample_factor = int(sampling_rate // rate)
y_signal = decimate(y_signal, down_sample_factor)
#x_signal = decimate(x_signal, down_sample_factor)
y_signal = normalize_data(pandas.DataFrame(y), max_range, min_range)
return x_signal, y_signal
def __call__(self, signal):
import biodata
from scipy.signal import find_peaks
from biosppy.signals.tools import get_heart_rate, smoother
list_smoothing_heart = [None, 10, 100]
list_smoothing_rate = [None, 1000, 10000]
list_smoothing_eda = [None, 10]
list_features = []
heart_raw = biodata.enveloppe_filter(signal[:, 1])
heart_peaks, heart_properties = find_peaks(
heart_raw,
distance=self.distance_heart,
width=self.width_heart,
prominence=self.prominence_heart)
for smoothing in list_smoothing_heart:
if smoothing is None:
size = 1
smoothing = False
else:
size = smoothing
smoothing = True
bpm = get_heart_rate(heart_peaks,
sampling_rate=self.sampling_rate,
smooth=smoothing,
size=size)
amplitudes = heart_properties["prominences"]
if smoothing:
amplitudes, _ = smoother(signal=amplitudes,
kernel='boxcar',
size=size,
mirror=True)
bpm = biodata.interpolate(bpm[1], bpm[0], len(signal))
amplitudes = biodata.interpolate(amplitudes, heart_peaks,
len(signal))
list_features += [bpm, amplitudes]
eda_raw = biodata.enveloppe_filter(signal[:, 2])
eda_peaks, eda_properties = find_peaks(eda_raw,
distance=self.distance_eda,
width=self.width_eda,
prominence=self.prominence_eda)
for smoothing in list_smoothing_eda:
if smoothing is None:
size = 1
smoothing = False
else:
size = smoothing
smoothing = True
intervals = biodata.compute_intervals(eda_peaks,
smooth=smoothing,
size=size)
amplitudes = eda_properties["prominences"]
if smoothing:
amplitudes, _ = smoother(signal=amplitudes,
kernel='boxcar',
size=size,
mirror=True)
intervals = biodata.interpolate(intervals, eda_peaks, len(signal))
amplitudes = biodata.interpolate(amplitudes, eda_peaks,
len(signal))
list_features += [intervals, amplitudes]
for smoothing in list_smoothing_rate:
if smoothing is None:
size = 1
smoothing = False
else:
size = smoothing
smoothing = True
rate = biodata.rate_of_change(eda_raw, size=size)
list_features.append(rate)
features = np.array(list_features).transpose()
if self.normalization is None:
pass
elif self.normalization == "standardized":
features = (features - features.mean(0)) / features.std(0)
elif self.normalization == "normalized":
features = (features - features.min(0)) / (features.max(0) -
features.min(0))
else:
raise ValueError
return torch.tensor(features).float()
def resp(signal=None, sampling_rate=1000., show=True):
"""Process a raw Respiration signal and extract relevant signal features
using default parameters.
Parameters
----------
signal : array
Raw Respiration signal.
sampling_rate : int, float, optional
Sampling frequency (Hz).
show : bool, optional
If True, show a summary plot.
Returns
-------
ts : array
Signal time axis reference (seconds).
filtered : array
Filtered Respiration signal.
zeros : array
Indices of Respiration zero crossings.
resp_rate_ts : array
Inspiration rate time axis reference (seconds).
resp_rate : array
Instantaneous respiration rate (Hz).
"""
# check inputs
if signal is None:
raise TypeError("Please specify an input signal.")
# ensure numpy
signal = np.array(signal)
sampling_rate = float(sampling_rate)
# filter signal
# 0.1 ~~ 0.35 Hzのバンドパスフィルタ
filtered, _, _ = st.filter_signal(signal=signal,
ftype='butter',
band='bandpass',
order=2,
frequency=[0.1, 0.35],
sampling_rate=sampling_rate)
# compute zero crossings
filtered = filtered - np.mean(filtered)
# zeros
df = np.diff(np.sign(filtered))
inspiration = np.nonzero(df > 0)[0]
expiration = np.nonzero(df < 0)[0]
if len(inspiration) < 2:
rate_idx = []
rate = []
else:
# compute resp peaks between inspiration and expiration
peaks = []
for i in range(len(inspiration) - 1):
cycle = filtered[inspiration[i]:inspiration[i + 1]]
peaks.append(np.argmax(cycle) + inspiration[i])
# list to array
peaks = np.array(peaks)
# compute respiration rate
rate_idx = inspiration[1:]
rate = sampling_rate * (1. / np.diff(inspiration))
# physiological limits
# 0.35Hz以下のresp_rateは省かれる
indx = np.nonzero(rate <= 0.35)
rate_idx = rate_idx[indx]
rate = rate[indx]
# smooth with moving average
size = 3
rate, _ = st.smoother(signal=rate,
kernel='boxcar',
size=size,
mirror=True)
# get time vectors
length = len(signal)
T = (length - 1) / sampling_rate
ts = np.linspace(0, T, length, endpoint=True)
ts_rate = ts[rate_idx]
# plot
if show:
plotting.plot_resp(ts=ts,
raw=signal,
filtered=filtered,
zeros=zeros,
resp_rate_ts=ts_rate,
resp_rate=rate,
path=None,
show=True)
# output
args = (ts, filtered, ts_rate, rate, inspiration, expiration, peaks)
names = ('ts', 'filtered', 'resp_rate_ts', 'resp_rate', 'inspiration',
'expiration', 'peaks')
return utils.ReturnTuple(args, names)
def ecg(
Kb=130,
Ap=0.2,
Kp=100,
Kpq=40,
Aq=0.1,
Kq1=25,
Kq2=5,
Ar=0.7,
Kr=40,
As=0.2,
Ks=30,
Kcs=5,
sm=96,
Kst=100,
At=0.15,
Kt=220,
si=2,
Ki=200,
var=0.01,
sampling_rate=10000,
): # normal values by default
"""Concatenates the segments and waves to make an ECG signal. The default values are physiological.
Follows the approach by Dolinský, Andráš, Michaeli and Grimaldi [Model03].
If the parameters introduced aren't within physiological values (limits based on the website [ECGwaves]), a warning will raise.
Parameters
----------
Kb : int, optional
B segment width (miliseconds).
Ap : float, optional
P wave amplitude (milivolts).
Kp : int, optional
P wave width (miliseconds).
Kpq : int, optional
PQ segment width (miliseconds).
Aq : float, optional
Q wave amplitude (milivolts).
Kq1 : int, optional
First 5/6 of the Q wave width (miliseconds).
Kq2 : int, optional
Last 1/6 of the Q wave width (miliseconds).
Ar : float, optional
R wave amplitude (milivolts).
Kr : int, optional
R wave width (miliseconds).
As : float, optional
S wave amplitude (milivolts).
Ks : int, optional
S wave width (miliseconds).
Kcs : int, optional
Parameter which allows slight adjustment of S wave shape by cutting away a portion at the end.
sm : int, optional
Slope parameter in the ST segment.
Kst : int, optional
ST segment width (miliseconds).
At : float, optional
1/2 of the T wave amplitude (milivolts).
Kt : int, optional
T wave width (miliseconds).
si : int, optional
Parameter for setting the transition slope between T wave and isoelectric line.
Ki : int, optional
I segment width (miliseconds).
var : float, optional
Value between 0.0 and 1.0 that adds variability to the obtained signal, by changing each parameter following a normal distribution with mean value `parameter_value` and std `var * parameter_value`.
sampling_rate : int, optional
Sampling frequency (Hz).
Returns
-------
ecg : array
Amplitude values of the ECG wave.
t : array
Time values accoring to the provided sampling rate.
params : dict
Input parameters of the function
Example
-------
sampling_rate = 10000
beats = 3
noise_amplitude = 0.05
ECGtotal = np.array([])
for i in range(beats):
ECGwave = ecg(sampling_rate=sampling_rate, var=0.1)
ECGtotal = np.concatenate((ECGtotal, ECGwave))
t = np.arange(0, len(ECGtotal)) / sampling_rate
# add powerline noise (50 Hz)
noise = noise_amplitude * np.sin(50 * (2 * pi) * t)
ECGtotal += noise
plt.plot(t, ECGtotal)
plt.xlabel("Time (ms)")
plt.ylabel("Amplitude (mV)")
plt.grid()
plt.title("ECG")
plt.show()
References
----------
.. [Model03] Pavol DOLINSKÝ, Imrich ANDRÁŠ, Linus MICHAELI, Domenico GRIMALDI,
"MODEL FOR GENERATING SIMPLE SYNTHETIC ECG SIGNALS",
Acta Electrotechnica et Informatica, Vol. 18, No. 3, 2018, 3–8
.. [ECGwaves] https://ecgwaves.com/
"""
if Kp > 120 and Ap >= 0.25:
warnings.warn("P wave isn't within physiological values.")
if Kq1 + Kq2 > 30 or Aq > 0.25 * Ar:
warnings.warn("Q wave isn't within physiological values.")
if 120 > Kp + Kpq or Kp + Kpq > 220:
warnings.warn("PR interval isn't within physiological limits.")
if Kq1 + Kq2 + Kr + Ks - Kcs > 120:
warnings.warn(
"QRS complex duration isn't within physiological limits.")
if Kq1 + Kq2 + Kr + Ks - Kcs + Kst + Kt > 450:
warnings.warn(
"QT segment duration isn't within physiological limits for men.")
if Kq1 + Kq2 + Kr + Ks - Kcs + Kst + Kt > 470:
warnings.warn(
"QT segment duration isn't within physiological limits for women.")
if var < 0 or var > 1:
raise TypeError("Variability value should be between 0.0 and 1.0")
if var > 0:
# change the parameter according to the provided variability
nd = lambda x: np.random.normal(x, x * var)
Kb = round(np.clip(nd(Kb), 0, 130))
Ap = np.clip(nd(Ap), -0.2, 0.5)
Kp = np.clip(nd(Kp), 10, 100)
Kpq = round(np.clip(nd(Kpq), 0, 60))
Aq = np.clip(nd(Aq), 0, 0.5)
Kq1 = round(np.clip(nd(Kq1), 0, 70))
Kq2 = round(np.clip(nd(Kq2), 0, 50))
Ar = np.clip(nd(Ar), 0.5, 2)
Kr = round(np.clip(nd(Kr), 10, 150))
As = np.clip(nd(As), 0, 1)
Ks = round(np.clip(nd(Ks), 10, 200))
Kcs = round(np.clip(nd(Kcs), -5, 150))
sm = round(np.clip(nd(sm), 1, 150))
Kst = round(np.clip(nd(Kst), 0, 110))
At = np.clip(nd(At), -0.5, 1)
Kt = round(np.clip(nd(Kt), 50, 300))
si = round(np.clip(nd(si), 0, 50))
# variable i is the time between samples (in miliseconds)
i = 1000 / sampling_rate
l = int(1 / i)
B_to_S = (B(Kb, l) + P(Ap, Kp, i) + Pq(Kpq, l) + Q1(Aq, Kq1, i) +
Q2(Aq, Kq2, i) + R(Ar, Kr, i) + S(As, Ks, Kcs, i))
St_to_I = (St(As, Ks, Kcs, sm, Kst, i) +
T(As, Ks, Kcs, sm, Kst, At, Kt, i) +
I(As, Ks, Kcs, sm, Kst, At, Kt, si, Ki, i))
# The signal is filtered in two different sizes
ECG1_filtered, n1 = st.smoother(B_to_S, size=50)
ECG2_filtered, n2 = st.smoother(St_to_I, size=500)
# The signal is concatenated
ECGwave = np.concatenate((ECG1_filtered, ECG2_filtered))
# Time array
t = np.arange(0, len(ECGwave)) / sampling_rate
# output
params = {
"Kb": 130,
"Ap": 0.2,
"Kp": 100,
"Kpq": 40,
"Aq": 0.1,
"Kq1": 25,
"Kq2": 5,
"Ar": 0.7,
"Kr": 40,
"As": 0.2,
"Ks": 30,
"Kcs": 5,
"sm": 96,
"Kst": 100,
"At": 0.15,
"Kt": 220,
"si": 2,
"Ki": 200,
"var": 0.01,
"sampling_rate": 10000,
}
args = (ECGwave, t, params)
names = ("ecg", "t", "params")
return utils.ReturnTuple(args, names)
band='bandpass',
order=order,
frequency=[2, 50],
sampling_rate=sampling_rate)
filtered_data_ch2, _, _ = st.filter_signal(signal=data_ch2_arr,
ftype='FIR',
band='bandpass',
order=order,
frequency=[2, 50],
sampling_rate=sampling_rate)
# Smooth
filtered_data_ch1, _ = st.smoother(signal=filtered_data_ch1,
kernel='hamming',
size=sm_size, mirror=True)
filtered_data_ch2, _ = st.smoother(signal=filtered_data_ch2,
kernel='hamming',
size=sm_size, mirror=True)
# Peaks
peaks_ch1 = find_peaks(filtered_data_ch1, sampling_rate)
peaks_ch2 = find_peaks(filtered_data_ch2, sampling_rate)
for i in range(0, len(peaks_ch1)):
peak1 = -5
def calculateECGFeatures(dataFrame, smooth=True, normalize=True):
'''
Inputs: dataFrame pandas.dataFrame containing clip of raw ECG data, should have two columns of
Timestamps (ms) and Sample (V)
smooth optional Boolean, default value is True, flag for whether to smooth input raw
ECG data or not
normalize option Boolean, default value is True, flag for whether to normalize input raw
ECG data or not
Outputs: features pandas.dataFrame for input clip of data, each column corresponds to a different
feature, included features are as follows:
- mean - ultra low frequency power
- median - very low frequency power
- max - low frequency power
- variance - high frequency power
- standard deviation - LF/HF ratio
- absolute deviation - total power
- kurtois - R-R interval
- skew
'''
# ----------------------------------------------------------------------------------------------------
# Data Preprocessing ---------------------------------------------------------------------------------
# ----------------------------------------------------------------------------------------------------
time = dataFrame['Timestamp (ms)'].values
data = dataFrame['Sample (V)'].values
## Smooth raw ECG data
if smooth:
smooth_signal = tools.smoother(
data, kernel='median', size=5) # Convolutional 5x5 kernel window
data = smooth_signal['signal']
## Normalize raw ECG data
if normalize:
norm_signal = tools.normalize(data)
data = norm_signal['signal']
# ----------------------------------------------------------------------------------------------------
# Begin Features Extraction Code ---------------------------------------------------------------------
# ----------------------------------------------------------------------------------------------------
## Calculate basic statistic features
s_mean, s_med, s_max, s_var, s_std_dev, s_abs_dev, s_kurtois, s_skew = tools.signal_stats(
data)
## Obtain Power Spectra
power_freqs, power_spectrum = tools.power_spectrum(data,
sampling_rate=2.0,
decibel=False)
## Calculate Ultra Low-Frequency Power (ULF Band: <= 0.003 Hz)
# power_ulf = tools.band_power(freqs=power_freqs, power=power_spectrum, frequency=[0, 0.003])
# power_ulf = np.absolute(power_ulf['avg_power'])
# Calculate Very Low-Frequency Power (VLF Band: 0.0033 - 0.04 Hz)
power_vlf = tools.band_power(freqs=power_freqs,
power=power_spectrum,
frequency=[0.0033, 0.04])
power_vlf = np.absolute(power_vlf['avg_power'])
# Calculate Low-Frequency Power (LF Band: 0.04 - 0.15 Hz Hz)
power_lf = tools.band_power(freqs=power_freqs,
power=power_spectrum,
frequency=[0.04, 0.15])
power_lf = np.absolute(power_lf['avg_power'])
## Calculate High-Frequency Power (HF Band: 0.15 - 0.40 Hz)
power_hf = tools.band_power(freqs=power_freqs,
power=power_spectrum,
frequency=[0.15, 0.40])
power_hf = np.absolute(power_hf['avg_power'])
## Calculate LF/HF Ratio
lf_hf_ratio = power_lf / power_hf
## Calculate Total Power (VLF + LF + HF power)
power_total = power_vlf + power_lf + power_hf
# Calculate R peak indices
r_peaks = ecg.engzee_segmenter(data, sampling_rate=500.0)
if np.size(r_peaks) != 1:
rr_indices = time[r_peaks]
rr_intervals = np.mean(np.diff(rr_indices))
else:
rr_intervals = np.nan # NaN indicates not enough R peaks within window to calculate R-R interval
# ----------------------------------------------------------------------------------------------------
# Collect Features and convert into dataFrame --------------------------------------------------------
# ----------------------------------------------------------------------------------------------------
signal_features = {
'mean': s_mean,
'median': s_med,
'max': s_max,
'variance': s_var,
'std_dev': s_std_dev,
'abs_dev': s_abs_dev,
'kurtois': s_kurtois,
'skew': s_skew,
# 'ulf_power': power_ulf,
'vlf_power': power_vlf,
'lf_power': power_lf,
'hf_power': power_hf,
'lf_hf_ratio': lf_hf_ratio,
'rr_interval': rr_intervals,
}
features = pd.Series(signal_features)
a = [
s_mean,
s_med,
s_max,
s_var,
s_std_dev,
s_abs_dev,
s_kurtois,
s_skew,
# power_ulf,
power_vlf,
power_lf,
power_hf,
lf_hf_ratio,
rr_intervals
# rr_indices
]
return a
