def test_ecg_clean():
sampling_rate = 1000
noise = 0.05
ecg = nk.ecg_simulate(sampling_rate=sampling_rate, noise=noise)
ecg_cleaned_nk = nk.ecg_clean(ecg,
sampling_rate=sampling_rate,
method="neurokit")
assert ecg.size == ecg_cleaned_nk.size
# Assert that highpass filter with .5 Hz lowcut was applied.
fft_raw = np.abs(np.fft.rfft(ecg))
fft_nk = np.abs(np.fft.rfft(ecg_cleaned_nk))
freqs = np.fft.rfftfreq(ecg.size, 1 / sampling_rate)
assert np.sum(fft_raw[freqs < .5]) > np.sum(fft_nk[freqs < .5])
# Comparison to biosppy (https://github.com/PIA-Group/BioSPPy/blob/e65da30f6379852ecb98f8e2e0c9b4b5175416c3/biosppy/signals/ecg.py#L69)
ecg_biosppy = nk.ecg_clean(ecg,
sampling_rate=sampling_rate,
method="biosppy")
original, _, _ = biosppy.tools.filter_signal(signal=ecg,
ftype='FIR',
band='bandpass',
order=int(0.3 *
sampling_rate),
frequency=[3, 45],
sampling_rate=sampling_rate)
assert np.allclose((ecg_biosppy - original).mean(), 0, atol=1e-6)
def extract_rpeak_features(row, signal):
"""
Extract the R peak features.
:param row: a `BaseDataset` row to calculate the features from
:param signal: the raw ECG signal
:return: `row` with the added features
"""
ecg_cleaned = nk.ecg_clean(signal, sampling_rate=row.Fs)
peaks, info = nk.ecg_peaks(ecg_cleaned, sampling_rate=row.Fs)
r_peaks_sec = np.where(peaks['ECG_R_Peaks'].to_numpy() == 1)[0].astype(
np.float32)
r_peaks_sec /= row.Fs # get R-peak times in seconds
num_peaks = len(r_peaks_sec)
if num_peaks > 2:
hrv = nk.hrv(peaks, sampling_rate=row.Fs, show=False).iloc[0]
row = row.append(hrv)
row['N_QRS'] = num_peaks
rr = np.diff(r_peaks_sec)
row = row.append(get_statistics(rr, 'RR'))
row = row.append(get_statistics(signal, 'signal'))
return row, info
def get_HRVs_values(data, header_data):
filter_lowcut = 0.001
filter_highcut = 15.0
filter_order = 1
tmp_hea = header_data[0].split(' ')
ptID = tmp_hea[0]
num_leads = int(tmp_hea[1])
sample_Fs = int(tmp_hea[2])
gain_lead = np.zeros(num_leads)
for ii in range(num_leads):
tmp_hea = header_data[ii + 1].split(' ')
gain_lead[ii] = int(tmp_hea[2].split('/')[0])
# for testing, we included the mean age of 57 if the age is a NaN
# This value will change as more data is being released
for iline in header_data:
if iline.startswith('#Age'):
tmp_age = iline.split(': ')[1].strip()
age = int(tmp_age if tmp_age != 'NaN' else 57)
elif iline.startswith('#Sex'):
tmp_sex = iline.split(': ')[1]
if tmp_sex.strip() == 'Female':
sex = 1
else:
sex = 0
elif iline.startswith('#Dx'):
label = iline.split(': ')[1].split(',')[0]
signal = data[1]
gain = gain_lead[1]
ecg_signal = nk.ecg_clean(signal * gain,
sampling_rate=sample_Fs,
method="biosppy")
_, rpeaks = nk.ecg_peaks(ecg_signal, sampling_rate=sample_Fs)
hrv_time = nk.hrv_time(rpeaks, sampling_rate=sample_Fs)
# hrv_non = nk.hrv_nonlinear(rpeaks, sampling_rate=sample_Fs)
try:
signal_peak, waves_peak = nk.ecg_delineate(ecg_signal,
rpeaks,
sampling_rate=sample_Fs)
p_peaks = waves_peak['ECG_P_Peaks']
except ValueError:
print('Exception raised!')
pass
p_peaks = np.asarray(p_peaks, dtype=float)
p_peaks = p_peaks[~np.isnan(p_peaks)]
p_peaks = [int(a) for a in p_peaks]
mean_P_Peaks = np.mean([signal[w] for w in p_peaks])
hrv_time['mean_P_Peaks'] = mean_P_Peaks
hrv_time['age'] = age
hrv_time['label'] = label
# df = pd.concat([hrv_time, hrv_non], axis=1)
return hrv_time
def test_ecg_rate():
sampling_rate = 1000
noise = 0.15
ecg = nk.ecg_simulate(duration=120,
sampling_rate=sampling_rate,
noise=noise,
random_state=42)
ecg_cleaned_nk = nk.ecg_clean(ecg,
sampling_rate=sampling_rate,
method="neurokit")
signals, info = nk.ecg_peaks(ecg_cleaned_nk, method="neurokit")
# Test without desired length.
rate = nk.ecg_rate(rpeaks=info, sampling_rate=sampling_rate)
assert rate.shape == (info["ECG_R_Peaks"].size, )
assert np.allclose(rate.mean(), 70, atol=2)
# Test with desired length.
test_length = 1200
rate = nk.ecg_rate(rpeaks=info,
sampling_rate=sampling_rate,
desired_length=test_length)
assert rate.shape == (test_length, )
assert np.allclose(rate.mean(), 70, atol=2)
def test_ecg_peaks():
sampling_rate = 1000
noise = 0.15
ecg = nk.ecg_simulate(duration=120,
sampling_rate=sampling_rate,
noise=noise,
random_state=42)
ecg_cleaned_nk = nk.ecg_clean(ecg,
sampling_rate=sampling_rate,
method="neurokit")
# Test without request to correct artifacts.
signals, info = nk.ecg_peaks(ecg_cleaned_nk,
correct_artifacts=False,
method="neurokit")
assert signals.shape == (120000, 1)
assert np.allclose(signals["ECG_R_Peaks"].values.sum(dtype=np.int64),
139,
atol=1)
# Test with request to correct artifacts.
signals, info = nk.ecg_peaks(ecg_cleaned_nk,
correct_artifacts=True,
method="neurokit")
assert signals.shape == (120000, 1)
assert np.allclose(signals["ECG_R_Peaks"].values.sum(dtype=np.int64),
139,
atol=1)
def qrs_detection_pantompkins_vs_neurokit(self, dataset):
row = dataset.data.iloc[50]
signal = dataset.read_record(row.Record)[:row.Fs * 20 + 1]
method_names = {'pantompkins': 'Pan–Tompkins', 'neurokit': 'Neurokit'}
for method in ['pantompkins', 'neurokit']:
ecg_cleaned = nk.ecg_clean(signal,
sampling_rate=row.Fs,
method=method)
peaks, info = nk.ecg_peaks(ecg_cleaned,
sampling_rate=row.Fs,
method=method)
r_peaks = np.where(peaks['ECG_R_Peaks'].to_numpy() == 1)[0]
fig = go.Figure()
fig.add_trace(
go.Scatter(x=np.arange(len(signal)) / row.Fs, y=signal))
fig.add_trace(
go.Scatter(mode='markers',
x=r_peaks / row.Fs,
y=signal[r_peaks]))
fig.update_traces(marker=dict(size=8))
self.set_ecg_layout(
fig,
title=f'{row.Record} - R peaks ({method_names[method]} method)',
showlegend=False,
xaxis=dict(range=[0, 20]),
yaxis=dict(range=[-5000, 5000]))
self.save_image(fig, f'qrs_{method}.png', width=900, height=300)
def process_X_values(X, Y):
dfs = []
Y["diagnostic_superclass"] = Y["diagnostic_superclass"].swifter.apply(
lambda x: 0 if x == "NORM" else
(1 if x == "MI" else (2 if x == "STTC" else (3 if x == "HYP" else 4))))
for v in tqdm(range(0, len(X))):
temp = pd.DataFrame(X[v],
columns=[
'I', 'II', 'III', 'aVL', 'aVR', 'aVF', 'V1',
'V2', 'V3', 'V4', 'V5', 'V6'
])
for value in temp.columns:
ecg = np.array(temp[value])
try:
signals = nk.ecg_clean(ecg,
sampling_rate=250,
method='pantompkins1985')
except:
signals = nk.ecg_clean(ecg,
sampling_rate=100,
method='pantompkins1985')
temp[value] = signals
temp['id'] = Y.iloc[v].patient_id
s = temp.groupby('id').cumcount().add(1)
temp = (temp.set_index(['id', s]).unstack().sort_index(axis=1,
level=1))
temp['diagnostic_superclass'] = Y.iloc[v].diagnostic_superclass
temp['strat_fold'] = Y.iloc[v].strat_fold
temp['id'] = Y.iloc[v].patient_id
dfs.append(temp)
data = pd.concat(dfs)
data = data[~np.isnan(data.id)]
return data
def apply(sampling_rate):
global x_global, y_global
ecg = []
ecg.append(x_global)
ecg.append(y_global)
window_autofilter.destroy()
ecg[1] = nk.ecg_clean(ecg[1], sampling_rate = sampling_rate)
x_global = ecg[0]
y_global = ecg[1]
a.clear()
a.plot(ecg[0], ecg[1])
plt.xlabel('time [s]')
plt.ylabel('voltage [mV]')
canvas.draw()
def test_ecg_findpeaks():
sampling_rate = 1000
ecg = nk.ecg_simulate(duration=60, sampling_rate=sampling_rate, noise=0, method="simple", random_state=42)
ecg_cleaned = nk.ecg_clean(ecg, sampling_rate=sampling_rate, method="neurokit")
# Test neurokit methodwith show=True
info_nk = nk.ecg_findpeaks(ecg_cleaned, show=True)
assert info_nk["ECG_R_Peaks"].size == 69
# This will identify the latest figure.
fig = plt.gcf()
assert len(fig.axes) == 2
# Test pantompkins1985 method
info_pantom = nk.ecg_findpeaks(nk.ecg_clean(ecg, method="pantompkins1985"), method="pantompkins1985")
assert info_pantom["ECG_R_Peaks"].size == 70
# Test hamilton2002 method
info_hamilton = nk.ecg_findpeaks(nk.ecg_clean(ecg, method="hamilton2002"), method="hamilton2002")
assert info_hamilton["ECG_R_Peaks"].size == 69
# Test christov2004 method
info_christov = nk.ecg_findpeaks(ecg_cleaned, method="christov2004")
assert info_christov["ECG_R_Peaks"].size == 273
# Test gamboa2008 method
info_gamboa = nk.ecg_findpeaks(ecg_cleaned, method="gamboa2008")
assert info_gamboa["ECG_R_Peaks"].size == 69
# Test elgendi2010 method
info_elgendi = nk.ecg_findpeaks(nk.ecg_clean(ecg, method="elgendi2010"), method="elgendi2010")
assert info_elgendi["ECG_R_Peaks"].size == 70
# Test engzeemod2012 method
info_engzeemod = nk.ecg_findpeaks(nk.ecg_clean(ecg, method="engzeemod2012"), method="engzeemod2012")
assert info_engzeemod["ECG_R_Peaks"].size == 70
# Test kalidas2017 method
info_kalidas = nk.ecg_findpeaks(nk.ecg_clean(ecg, method="kalidas2017"), method="kalidas2017")
assert np.allclose(info_kalidas["ECG_R_Peaks"].size, 68, atol=1)
# Test martinez2003 method
ecg = nk.ecg_simulate(duration=60, sampling_rate=sampling_rate, noise=0, random_state=42)
ecg_cleaned = nk.ecg_clean(ecg, sampling_rate=sampling_rate, method="neurokit")
info_martinez = nk.ecg_findpeaks(ecg_cleaned, method="martinez2003")
assert np.allclose(info_martinez["ECG_R_Peaks"].size, 69, atol=1)
def generate_features(ecg, header):
#input: 12-lead ecg and its header
fs = 500
features = {}
lead_names = []
for iline in header:
if '.mat' in iline:
name = iline.split(' 0 ')[2].strip()
lead_names.append(name)
for ecg_signal, lead in zip(ecg, lead_names):
ecg_cleaned = nk.ecg_clean(ecg_signal, sampling_rate=fs)
if np.all((ecg_cleaned == 0)):
return None
else:
_, rpeaks = nk.ecg_peaks(ecg_cleaned, sampling_rate=fs)
if rpeaks['ECG_R_Peaks'].size == 0:
return None
else:
try:
signal_dwt, waves_dwt = nk.ecg_delineate(
ecg_cleaned,
rpeaks['ECG_R_Peaks'],
sampling_rate=fs,
method="dwt")
biphase, areas, t_till_peaks, ampls, dur, idxs, pq_intervals = p_peak_features(
ecg_cleaned, waves_dwt)
features_for_single_lead = {
'PQ_int': calculate_features(pq_intervals),
'P_dur': calculate_features(dur),
'Area/Dur_P': calculate_features(idxs),
'Area_under_P': calculate_features(areas),
'P_amp': calculate_features(ampls),
'Time_till_P': calculate_features(t_till_peaks),
'Biphase_P': calculate_features(biphase)
}
except IndexError:
return None
features[lead] = features_for_single_lead
return features
def my_processing(ecg_signal):
# Try processing
ecg_cleaned = nk.ecg_clean(ecg_signal, sampling_rate=300, method="biosppy")
instant_peaks, rpeaks = nk.ecg_peaks(ecg_cleaned,
sampling_rate=300,
method='hamilton2002')
info = rpeaks
try:
# Additional info of the ecg signal
delineate_signal, delineate_waves = nk.ecg_delineate(
ecg_cleaned=ecg_cleaned,
rpeaks=rpeaks,
sampling_rate=300,
method='cwt')
except:
delineate_signal = np.NaN
delineate_waves = np.NaN
return ecg_cleaned, delineate_signal, delineate_waves, info
def find_R_peaks(ecg_data, samplefreq):
try:
_, rpeaks = nk.ecg_peaks(ecg_data, sampling_rate=samplefreq)
r_peaks = rpeaks['ECG_R_Peaks']
r_peaks = np.delete(r_peaks,
np.where(np.isnan(r_peaks))[0]).astype(int)
except:
print("cleaning data")
cleaned_ecg = nk.ecg_clean(ecg_data,
sampling_rate=samplefreq,
method="neurokit")
try:
_, rpeaks = nk.ecg_peaks(cleaned_ecg, sampling_rate=samplefreq)
r_peaks = rpeaks['ECG_R_Peaks']
r_peaks = np.delete(r_peaks,
np.where(np.isnan(r_peaks))[0]).astype(int)
except:
print("could not analyse cleaned ECG")
#Midlertidig løsning:
r_peaks = np.array([0, 1, 2, 3])
return r_peaks
def extract_features_tmaps(
self,
signal_tm: TensorMap,
clean_method: str = "neurokit",
r_method: str = "neurokit",
wave_method: str = "dwt",
min_peaks: int = 200,
):
"""
Function to extract the ecg features using the neurokit2 package. That
is the P, Q, R, S and T peaks and the P, QRS and T waves onsets and
offsets. The result is saved internally.
:param signal_tm: <TensorMap>
:param clean_method: <str> The processing pipeline to apply. Can be one of
‘neurokit’ (default), ‘biosppy’, ‘pantompkins1985’,
‘hamilton2002’, ‘elgendi2010’, ‘engzeemod2012’.
:param r_method: <str> The algorithm to be used for R-peak detection. Can be one
of ‘neurokit’ (default), ‘pantompkins1985’, ‘hamilton2002’,
‘christov2004’, ‘gamboa2008’, ‘elgendi2010’, ‘engzeemod2012’
or ‘kalidas2017’.
:param wave_method: <str> Can be one of ‘dwt’ (default) for discrete
wavelet transform or ‘cwt’ for continuous wavelet transform.
:param min_peaks: <int> Minimum R peaks to be detected to proceed with
further calculations.
"""
for i, _ in enumerate(self.sampling_rate):
sampling_rate = self.sampling_rate[i][0]
init = self.sampling_rate[i][1]
if i == len(self.sampling_rate) - 1:
end = -1
else:
end = self.sampling_rate[i + 1][1]
ecg_signal = signal_tm.tensor_from_file(signal_tm,
self)[0][init:end]
ecg_signal = nk.ecg_clean(ecg_signal, sampling_rate, clean_method)
try:
_, r_peaks = nk.ecg_peaks(ecg_signal, sampling_rate, r_method)
except IndexError:
continue
if len(r_peaks["ECG_R_Peaks"]) < min_peaks:
continue
_, waves_peaks = nk.ecg_delineate(ecg_signal, r_peaks,
sampling_rate)
_, waves_peaks_2 = nk.ecg_delineate(
ecg_signal,
r_peaks,
sampling_rate,
wave_method,
)
waves_peaks.update(waves_peaks_2)
for peak_type in r_peaks:
if peak_type not in self.r_peaks:
self.r_peaks[peak_type] = r_peaks[peak_type]
else:
self.r_peaks[peak_type] = np.append(
self.r_peaks[peak_type],
r_peaks[peak_type],
)
for peak_type in waves_peaks:
if peak_type not in self.waves_peaks:
self.waves_peaks[peak_type] = waves_peaks[peak_type]
else:
self.waves_peaks[peak_type] = np.append(
self.waves_peaks[peak_type],
waves_peaks[peak_type],
)
for peak_type in self.r_peaks:
self.r_peaks[peak_type] = list(self.r_peaks[peak_type])
for peak_type in self.waves_peaks:
self.waves_peaks[peak_type] = list(self.waves_peaks[peak_type])
def extract_features(
self,
clean_method: str = "neurokit",
r_method: str = "neurokit",
wave_method: str = "dwt",
min_peaks: int = 200,
size: int = 200000,
):
"""
Function to extract the ecg features using the neurokit2 package. That
is the P, Q, R, S and T peaks and the P, QRS and T waves onsets and
offsets. The result is saved internally.
:param clean_method: <str> The processing pipeline to apply. Can be one of
‘neurokit’ (default), ‘biosppy’, ‘pantompkins1985’,
‘hamilton2002’, ‘elgendi2010’, ‘engzeemod2012’.
:param r_method: <str> The algorithm to be used for R-peak detection. Can be one
of ‘neurokit’ (default), ‘pantompkins1985’, ‘hamilton2002’,
‘christov2004’, ‘gamboa2008’, ‘elgendi2010’, ‘engzeemod2012’
or ‘kalidas2017’.
:param wave_method: <str> Can be one of ‘dwt’ (default) for discrete
wavelet transform or ‘cwt’ for continuous wavelet transform.
:param min_peaks: <int> Minimum R peaks to be detected to proceed with
further calculations.
:param size: <int> ECG sample size to analyze per loop.
"""
if not self.lead:
return
for i, _ in enumerate(self.sampling_rate):
sampling_rate = self.sampling_rate[i][0]
init = self.sampling_rate[i][1]
if i == len(self.sampling_rate) - 1:
ecg_signal_size = (
ECG_TMAPS[f"{self.lead}_value"].tensor_from_file(
ECG_TMAPS[f"{self.lead}_value"],
self,
visit=self.visit,
)[0][init:].shape[0])
else:
ecg_signal_size = self.sampling_rate[i + 1][1] - init
if size < ecg_signal_size:
end = init + size
else:
end = init + ecg_signal_size
while init < ecg_signal_size + self.sampling_rate[i][1]:
ecg_signal = ECG_TMAPS[f"{self.lead}_value"].tensor_from_file(
ECG_TMAPS[f"{self.lead}_value"],
self,
visit=self.visit,
)[0][init:end]
ecg_signal = nk.ecg_clean(ecg_signal, sampling_rate,
clean_method)
try:
_, r_peaks = nk.ecg_peaks(ecg_signal, sampling_rate,
r_method)
except IndexError:
init = end
end = init + size
if end > ecg_signal_size + self.sampling_rate[i][1]:
end = ecg_signal_size + self.sampling_rate[i][1]
continue
if len(r_peaks["ECG_R_Peaks"]) < min_peaks:
init = end
end = init + size
if end > ecg_signal_size + self.sampling_rate[i][1]:
end = ecg_signal_size + self.sampling_rate[i][1]
continue
_, waves_peaks = nk.ecg_delineate(ecg_signal, r_peaks,
sampling_rate)
_, waves_peaks_2 = nk.ecg_delineate(
ecg_signal,
r_peaks,
sampling_rate,
wave_method,
)
waves_peaks.update(waves_peaks_2)
for peak_type in r_peaks:
if peak_type not in self.r_peaks:
self.r_peaks[peak_type] = r_peaks[peak_type]
else:
self.r_peaks[peak_type] = np.append(
self.r_peaks[peak_type],
r_peaks[peak_type],
)
for peak_type in waves_peaks:
if peak_type not in self.waves_peaks:
self.waves_peaks[peak_type] = waves_peaks[peak_type]
else:
self.waves_peaks[peak_type] = np.append(
self.waves_peaks[peak_type],
waves_peaks[peak_type],
)
init = end
end = init + size
if end > ecg_signal_size + self.sampling_rate[i][1]:
end = ecg_signal_size + self.sampling_rate[i][1]
for peak_type in self.r_peaks:
self.r_peaks[peak_type] = list(self.r_peaks[peak_type])
for peak_type in self.waves_peaks:
self.waves_peaks[peak_type] = list(self.waves_peaks[peak_type])
def create_df(dataframe: pd.DataFrame) -> pd.DataFrame:
# get lengths of signals for each sample
lengths = []
width = dataframe.shape[1]
for row in dataframe.index.tolist():
temp_width = width
for item in dataframe.loc[row][::-1]:
if not pd.isna(item) and isinstance(item, float):
temp_width -= 1
break
temp_width -= 1
lengths.append(temp_width)
"""
README
For the following features we measured: [mean, median, 5 % percentile, 95 % percentile, standard deviation]
R-peak location were retrieved by nk.ecg_peaks
Q-peak and S-location were retrieved by nk.ecg_delineate
?_ampl_* ?-Peak amplitude
?_nr_peaks number of ?-Peaks
?_diff_* Interval between ?-Peaks
QRS_diff_* QRS duration
len_* length of signal
Qual_* quality of signal measured with nk.ecg_quality
sign_* signal
Also the output from nk.hrv_time which contains different measurements for the heart rate variation (HRV*) was added
Additionally one 'typical' heartbeat was greated (all length 180):
MN_* mean signal
MD_* median signal
P5_* 5 % percentile signal
P95_* 95 % percentile signal
SD_* standard deviation of signal
"""
names = ['R_ampl_mean', 'R_ampl_median', 'R_ampl_perc5', 'R_ampl_perc95', 'R_ampl_sd', 'R_nr_peaks',
'len_mean', 'len_median', 'len_perc5', 'len_perc95', 'len_sd',
'sign_mean', 'sign_median', 'sign_perc5', 'sign_perc95', 'sign_sd',
'Qual_mean', 'Qual_median', 'Qual_perc5', 'Qual_perc95', 'Qual_sd',
'Q_ampl_mean', 'Q_ampl_median', 'Q_ampl_perc5', 'Q_ampl_perc95', 'Q_ampl_sd', 'Q_nr_peaks',
'Q_diff_mean', 'Q_diff_median', 'Q_diff_perc5', 'Q_diff_perc95', 'Q_diff_sd',
'S_ampl_mean', 'S_ampl_median', 'S_ampl_perc5', 'S_ampl_perc95', 'S_ampl_sd', 'S_nr_peaks',
'S_diff_mean', 'S_diff_median', 'S_diff_perc5', 'S_diff_perc95', 'S_diff_sd',
'P_ampl_mean', 'P_ampl_median', 'P_ampl_perc5', 'P_ampl_perc95', 'P_ampl_sd', 'P_nr_peaks',
'T_ampl_mean', 'T_ampl_median', 'T_ampl_perc5', 'T_ampl_perc95', 'T_ampl_sd', 'T_nr_peaks',
'QRS_diff_mean', 'QRS_diff_median', 'QRS_diff_perc5', 'QRS_diff_perc95', 'QRS_diff_sd',
'PR_diff_mean', 'PR_diff_median', 'PR_diff_perc5', 'PR_diff_perc95', 'PR_diff_sd',
'RT_diff_mean', 'RT_diff_median', 'RT_diff_perc5', 'RT_diff_perc95', 'RT_diff_sd',
'HRV_RMSSD', 'HRV_MeanNN', 'HRV_SDNN', 'HRV_SDSD', 'HRV_CVNN', 'HRV_CVSD', 'HRV_MedianNN',
'HRV_MadNN', 'HRV_MCVNN', 'HRV_IQRNN', 'HRV_pNN50', 'HRV_pNN20', 'HRV_TINN', 'HRV_HTI',
'HRV_ULF','HRV_VLF','HRV_LF','HRV_HF','HRV_VHF','HRV_LFHF','HRV_LFn','HRV_HFn', 'HRV_LnHF',
'HRV_SD1','HRV_SD2', 'HRV_SD1SD2','HRV_S','HRV_CSI','HRV_CVI','HRV_CSI_Modified', 'HRV_PIP',
'HRV_IALS','HRV_PSS','HRV_PAS','HRV_GI','HRV_SI','HRV_AI','HRV_PI','HRV_C1d','HRV_C1a','HRV_SD1d',
'HRV_SD1a','HRV_C2d','HRV_C2a','HRV_SD2d','HRV_SD2a','HRV_Cd','HRV_Ca','HRV_SDNNd','HRV_SDNNa','HRV_ApEn',
'HRV_SampEn','J_LF','J_HF','J_L/H']
template_len = 180
mean_names = ['MN_' + str(index) for index in range(template_len)]
median_names = ['MD_' + str(index) for index in range(template_len)]
perc5_names = ['P5_' + str(index) for index in range(template_len)]
perc95_names = ['P95_' + str(index) for index in range(template_len)]
sd_names = ['SD_' + str(index) for index in range(template_len)]
wavelet = 'db3'
wl_len = int(np.floor((template_len + pywt.Wavelet(wavelet).dec_len - 1) / 2))
wl_mean_names = ['WLMN_' + str(index) for index in range(2*wl_len)]
wl_median_names = ['WLMD_' + str(index) for index in range(2*wl_len)]
wl_perc5_names = ['WLP5_' + str(index) for index in range(2*wl_len)]
wl_perc95_names = ['WLP95_' + str(index) for index in range(2*wl_len)]
wl_sd_names = ['WLSD_' + str(index) for index in range(2*wl_len)]
typical_signal_names = mean_names + median_names + perc5_names + perc95_names + sd_names + wl_mean_names + \
wl_median_names + wl_perc5_names + wl_perc95_names + wl_sd_names
names += typical_signal_names
data = np.empty([dataframe.shape[0], len(names)])
iteration = 0
for row_index, row in dataframe.iterrows():
print(row_index)
# Retrieve ECG data
ecg_signal = row[:lengths[iteration] + 1]
ecg_signal = nk.ecg_clean(ecg_signal, sampling_rate=SAMPLING_RATE)
# Find R-peaks
peaks, info = nk.ecg_peaks(ecg_signal, sampling_rate=SAMPLING_RATE)
# R amplitude
R_amplitudes = ecg_signal[info['ECG_R_Peaks']]
# Check if the signal is flipped
# Check if we have enough peaks to retrieve more information
if len(R_amplitudes) > 4:
_, waves_peak = nk.ecg_delineate(ecg_signal, info, sampling_rate=300, show=False)
# Q amplitude
# remove nan values
Q_amplitudes = [ecg_signal[peak_index] if str(peak_index) != 'nan' else - np.infty for peak_index in
waves_peak['ECG_Q_Peaks']]
if np.sum([1 if np.abs(rpeak) > np.abs(Q_amplitudes[index]) else -1 for index, rpeak in
enumerate(R_amplitudes)]) < 0:
print("flip", row_index)
ecg_signal = -ecg_signal
peaks, info = nk.ecg_peaks(ecg_signal, sampling_rate=300)
# R amplitude
R_amplitudes = ecg_signal[info['ECG_R_Peaks']]
if len(R_amplitudes) > 4:
_, waves_peak = nk.ecg_delineate(ecg_signal, info, sampling_rate=300, show=False)
data_temp = []
if len(R_amplitudes) > 0:
data_temp = [np.mean(R_amplitudes),
np.median(R_amplitudes),
np.percentile(R_amplitudes, q=5),
np.percentile(R_amplitudes, q=95),
np.std(R_amplitudes),
len(R_amplitudes)]
else:
empty = np.empty([6])
empty[:] = np.NaN
data_temp += empty.tolist()
# length of signal
data_new = [np.mean(lengths[iteration] / SAMPLING_RATE),
np.median(lengths[iteration] / SAMPLING_RATE),
np.percentile(lengths[iteration] / SAMPLING_RATE, q=5),
np.percentile(lengths[iteration] / SAMPLING_RATE, q=95),
np.std(lengths[iteration] / SAMPLING_RATE)]
data_temp += data_new
# signal
data_new = [np.mean(ecg_signal),
np.median(ecg_signal),
np.percentile(ecg_signal, q=5),
np.percentile(ecg_signal, q=95),
np.std(ecg_signal)]
data_temp += data_new
# Check if we have enough peaks to retrieve more information
if len(R_amplitudes) > 4:
quality = nk.ecg_quality(ecg_signal, sampling_rate=SAMPLING_RATE)
data_new = [np.mean(quality),
np.median(quality),
np.percentile(quality, q=5),
np.percentile(quality, q=95),
np.std(quality)]
data_temp += data_new
# Delineate the ECG signal
# “ECG_P_Peaks”, “ECG_Q_Peaks”, “ECG_S_Peaks”, “ECG_T_Peaks”, “ECG_P_Onsets”, “ECG_T_Offsets”
# _, waves_peak = nk.ecg_delineate(ecg_signal, info, sampling_rate=SAMPLING_RATE, show=False)
# Q amplitude
# remove nan values
Q_peaks = [peak for peak in waves_peak['ECG_Q_Peaks'] if str(peak) != 'nan']
if len(Q_peaks) > 0:
Q_amplitudes = ecg_signal[Q_peaks]
data_new = [np.mean(Q_amplitudes),
np.median(Q_amplitudes),
np.percentile(Q_amplitudes, q=5),
np.percentile(Q_amplitudes, q=95),
np.std(Q_amplitudes),
len(Q_amplitudes)]
data_temp += data_new
else:
empty = np.empty([6])
empty[:] = np.NaN
empty[5] = 0
data_temp += empty.tolist()
# more than 1 Q-Peak => can build interval[s]
if len(Q_peaks) > 1:
Q_peaks_diff = [(Q_peaks[index + 1] - Q_peaks[index]) / SAMPLING_RATE
for index, item in enumerate(Q_peaks[:len(Q_peaks) - 1])]
# QQ interval
data_new = [np.mean(Q_peaks_diff),
np.median(Q_peaks_diff),
np.percentile(Q_peaks_diff, q=5),
np.percentile(Q_peaks_diff, q=95),
np.std(Q_peaks_diff)]
data_temp += data_new
# 0 or 1 Q-peak = no interval => return nan
else:
empty = np.empty([5])
empty[:] = np.NaN
data_temp += empty.tolist()
# S amplitude
# remove nan values
S_peaks = [peak for peak in waves_peak['ECG_S_Peaks'] if str(peak) != 'nan']
if len(S_peaks) > 0:
S_amplitudes = ecg_signal[S_peaks]
data_new = [np.mean(S_amplitudes),
np.median(S_amplitudes),
np.percentile(S_amplitudes, q=5),
np.percentile(S_amplitudes, q=95),
np.std(S_amplitudes),
len(S_amplitudes)]
data_temp += data_new
else:
empty = np.empty([6])
empty[:] = np.NaN
empty[5] = 0
data_temp += empty.tolist()
# more than one S-peak
if len(S_peaks) > 1:
S_peaks_diff = [(S_peaks[index + 1] - S_peaks[index]) / SAMPLING_RATE
for index, item in enumerate(S_peaks[:len(S_peaks) - 1])]
# SS interval
data_new = [np.mean(S_peaks_diff),
np.median(S_peaks_diff),
np.percentile(S_peaks_diff, q=5),
np.percentile(S_peaks_diff, q=95),
np.std(S_peaks_diff)]
data_temp += data_new
# 0 or 1 S-peak = no interval => return nan
else:
empty = np.empty([5])
empty[:] = np.NaN
data_temp += empty.tolist()
P_peaks = [peak for peak in waves_peak['ECG_P_Peaks'] if str(peak) != 'nan']
if len(P_peaks) > 0:
P_amplitudes = ecg_signal[P_peaks]
data_new = [np.mean(P_amplitudes),
np.median(P_amplitudes),
np.percentile(P_amplitudes, q=5),
np.percentile(P_amplitudes, q=95),
np.std(P_amplitudes),
len(P_amplitudes)]
data_temp += data_new
else:
empty = np.empty([6])
empty[:] = np.NaN
empty[5] = 0
data_temp += empty.tolist()
T_peaks = [peak for peak in waves_peak['ECG_T_Peaks'] if str(peak) != 'nan']
if len(T_peaks) > 0:
T_peaks = ecg_signal[T_peaks]
data_new = [np.mean(T_peaks),
np.median(T_peaks),
np.percentile(T_peaks, q=5),
np.percentile(T_peaks, q=95),
np.std(T_peaks),
len(T_peaks)]
data_temp += data_new
else:
empty = np.empty([6])
empty[:] = np.NaN
empty[5] = 0
data_temp += empty.tolist()
# QRS interval
QRS_peaks_diff = []
# compute difference between Q and S peak
for index in range(len(waves_peak['ECG_Q_Peaks'])):
if not (np.isnan(waves_peak['ECG_Q_Peaks'][index]) or np.isnan(waves_peak['ECG_S_Peaks'][index])):
QRS_peaks_diff.append(
(waves_peak['ECG_S_Peaks'][index] - waves_peak['ECG_Q_Peaks'][index]) / SAMPLING_RATE)
if len(QRS_peaks_diff) > 0:
data_new = [np.mean(QRS_peaks_diff),
np.median(QRS_peaks_diff),
np.percentile(QRS_peaks_diff, q=5),
np.percentile(QRS_peaks_diff, q=95),
np.std(QRS_peaks_diff)]
data_temp += data_new
else:
empty = np.empty([5])
empty[:] = np.NaN
data_temp += empty.tolist()
# PR interval
PR_peaks_diff = []
# compute difference between P and R peak
for index in range(len(waves_peak['ECG_P_Peaks'])):
if not np.isnan(waves_peak['ECG_P_Peaks'][index]):
PR_peaks_diff.append(
(info['ECG_R_Peaks'][index] - waves_peak['ECG_P_Peaks'][index]) / SAMPLING_RATE)
if len(PR_peaks_diff) > 0:
data_new = [np.mean(PR_peaks_diff),
np.median(PR_peaks_diff),
np.percentile(PR_peaks_diff, q=5),
np.percentile(PR_peaks_diff, q=95),
np.std(PR_peaks_diff)]
data_temp += data_new
else:
empty = np.empty([5])
empty[:] = np.NaN
data_temp += empty.tolist()
# RT interval
RT_peaks_diff = []
# compute difference between P and R peak
for index in range(len(waves_peak['ECG_T_Peaks'])):
if not np.isnan(waves_peak['ECG_T_Peaks'][index]):
RT_peaks_diff.append(
(waves_peak['ECG_T_Peaks'][index] - info['ECG_R_Peaks'][index]) / SAMPLING_RATE)
if len(RT_peaks_diff) > 0:
data_new = [np.mean(RT_peaks_diff),
np.median(PR_peaks_diff),
np.percentile(RT_peaks_diff, q=5),
np.percentile(RT_peaks_diff, q=95),
np.std(RT_peaks_diff)]
data_temp += data_new
else:
empty = np.empty([5])
empty[:] = np.NaN
data_temp += empty.tolist()
# Extract clean EDA and SCR features
# explanation of features:
# https://neurokit2.readthedocs.io/en/latest/functions.html?highlight=hrv%20time#neurokit2.hrv.hrv_time
hrv_time = nk.hrv(peaks, sampling_rate=SAMPLING_RATE, show=False)
data_new = hrv_time.values.tolist()[0]
data_temp += data_new
# Jannik
# http://www.paulvangent.com/2016/03/21/analyzing-a-discrete-heart-rate-signal-using-python-part-2/
rpeaks = info['ECG_R_Peaks']
r_interval = [rpeaks[index+1]-rpeaks[index] for index in range(len(rpeaks)-1)]
RR_x_new = np.linspace(rpeaks[0],rpeaks[-2],rpeaks[-2])
f = interp1d(rpeaks[:-1], r_interval, kind='cubic')
n = lengths[iteration] + 1 # Length of the signal
frq = np.fft.fftfreq(n, d=(1 / SAMPLING_RATE)) # divide the bins into frequency categories
frq = frq[range(int(n/2))] # Get single side of the frequency range
Y = np.fft.fft(f(RR_x_new))/n # Calculate FFT
try:
Y = Y[range(int(n / 2))]
lf = np.trapz(abs(Y[(frq >= 0.04) & (frq <= 0.15)]))
hf = np.trapz(abs(Y[(frq >= 0.16) & (frq <= 0.5)])) # Do the same for 0.16-0.5Hz (HF)
data_new = [lf, hf, lf / hf]
data_temp += data_new
except IndexError as err:
print(err)
data_temp += [None, None, None]
# if we don't have enough R peaks return vector of nan's
else:
empty = np.empty([len(names) - 16 - len(typical_signal_names)])
empty[:] = np.NaN
data_temp += empty.tolist()
# Create a 'typical' heartbeat
# Scaler = StandardScaler()
# ecg_signal = Scaler.fit_transform(X=ecg_signal.reshape(-1, 1)).reshape(1, -1)[0].tolist()
out = ecg.ecg(signal=ecg_signal, sampling_rate=SAMPLING_RATE, show=False)
mean = np.mean(out['templates'], axis=0)
median = np.median(out['templates'], axis=0)
perc5 = np.percentile(out['templates'].astype(np.float64), axis=0, q=5)
perc95 = np.percentile(out['templates'].astype(np.float64), axis=0, q=95)
std = np.std(out['templates'].astype(np.float64), axis=0)
data_new = np.concatenate((mean, median, perc5, perc95, std)).tolist()
data_temp += data_new
(wl_mean_cA, wl_mean_cD) = pywt.dwt(np.mean(out['templates'], axis=0),
'db3', 'periodic')
(wl_median_cA, wl_median_cD) = pywt.dwt(np.median(out['templates'], axis=0),
'db3', 'periodic')
(wl_perc5_cA, wl_perc5_cD) = pywt.dwt(np.percentile(out['templates'].astype(np.float64), axis=0, q=5),
'db3', 'periodic')
(wl_perc95_cA, wl_perc95_cD) = pywt.dwt(np.percentile(out['templates'].astype(np.float64), axis=0, q=95),
'db3', 'periodic')
(wl_sd_cA, wl_sd_cD) = pywt.dwt(np.std(out['templates'].astype(np.float64), axis=0),
'db3', 'periodic')
data_new = np.concatenate((wl_mean_cA, wl_mean_cD,
wl_median_cA, wl_median_cD,
wl_perc5_cA, wl_perc5_cD,
wl_perc95_cA, wl_perc95_cD,
wl_sd_cA, wl_sd_cD)).tolist()
data_temp += data_new
data[iteration] = data_temp
iteration += 1
features = pd.DataFrame(data, columns=names)
return features
def compute_features(data, condition, sampling_rate=700, window_size=60, window_shift=0.25):
index = 0
init = time.time()
# data cleaning
## ECG
ecg_cleaned = nk.ecg_clean(data["ECG"][condition].flatten(), sampling_rate=sampling_rate)
## == OLD
# ecg_rpeaks, _ = nk.ecg_peaks(ecg_cleaned, sampling_rate=sampling_rate)
# ecg_hr = nk.signal_rate(ecg_rpeaks, sampling_rate=sampling_rate)
## ==
## EDA
## 5Hz lowpass filter
eda_highcut = 5
eda_filtered = nk.signal_filter(data['EDA'][condition].flatten(), sampling_rate=sampling_rate, highcut=eda_highcut)
eda_cleaned = nk.standardize(eda_filtered)
# TODO: not sure about the approach. cvxeda takes longer periods
# phasic_tonic = nk.eda_phasic(cleaned, sampling_rate=700, method='cvxeda')
eda_phasic_tonic = nk.eda_phasic(eda_cleaned, sampling_rate=sampling_rate)
eda_phasic_tonic['t'] = [(1 / sampling_rate) * i for i in range(eda_phasic_tonic.shape[0])]
eda_scr_peaks, scr_info = nk.eda_peaks(eda_phasic_tonic['EDA_Phasic'], sampling_rate=sampling_rate)
## EMG
## For 5 sec window signal
## More on DC Bias https://www.c-motion.com/v3dwiki/index.php/EMG:_Removing_DC_Bias
emg_lowcut = 50
emg_filtered_dc = nk.signal_filter(data['EMG'][condition].flatten(), sampling_rate=sampling_rate, lowcut=emg_lowcut)
# OR 100 Hz highpass Butterworth filter followed by a constant detrending
# filtered_dc = nk.emg_clean(chest_data_dict['EMG'][baseline].flatten(), sampling_rate=700)
## For 60 sec window signal
# 50Hz lowpass filter
emg_highcut = 50
emg_filtered = nk.signal_filter(data['EMG'][condition].flatten(), sampling_rate=sampling_rate, highcut=emg_highcut)
## Resp
## Method biosppy important to appply bandpass filter 0.1 - 0.35 Hz
resp_processed, _ = nk.rsp_process(data['Resp'][condition].flatten(), sampling_rate=sampling_rate, method='biosppy')
print('Elapsed Preprocess', str(timedelta(seconds=time.time() - init)))
init = time.time()
chest_df_5 = pd.DataFrame() # For 5 sec window size
chest_df = pd.DataFrame()
window = int(sampling_rate * window_size)
for i in range(0, data['ACC'][condition].shape[0] - window, int(sampling_rate * window_shift)):
# ACC
w_acc_data = data['ACC'][condition][i: window + i]
acc_x_mean, acc_y_mean, acc_z_mean = np.mean(w_acc_data, axis=0) # Feature
acc_x_std, acc_y_std, acc_z_std = np.std(w_acc_data, axis=0) # Feature
acc_x_peak, acc_y_peak, acc_z_peak = np.amax(w_acc_data, axis=0) # Feature
acc_x_absint, acc_y_absint, acc_z_absint = np.abs(np.trapz(w_acc_data, axis=0)) # Feature
xyz = np.sum(w_acc_data, axis=0)
xyz_mean = np.mean(xyz) # Feature
xyz_std = np.std(xyz) # Feature
xyz_absint = np.abs(np.trapz(xyz)) # Feature
# == OLD
# ## ECG
# w_ecg_rpeaks = ecg_rpeaks[i: window + i]
# # HR
# w_ecg_hr = ecg_hr[i: window + i]
# hr_mean = np.mean(w_ecg_hr) # Feature
# hr_std = np.std(w_ecg_hr) # Feature
# # HRV Time-domain Indices
# # HRV_MeanNN
# # HRV_SDNN
# # HRV_pNN50
# # HRV_RMSSD -> Root mean square of the HRV
# # HRV_HTI -> Triangular interpolation index
# hrv_time = nk.hrv_time(w_ecg_rpeaks, sampling_rate=sampling_rate, show=False)
# hrv_mean = hrv_time.loc[0, 'HRV_MeanNN'] # Feature
# hrv_std = hrv_time.loc[0, 'HRV_SDNN'] # Feature
# # TODO: NN50
# # hrv_NN50 =
# hrv_pNN50 = hrv_time.loc[0, 'HRV_pNN50'] # Feature
# hrv_TINN = hrv_time.loc[0, 'HRV_HTI'] # Feature
# hrv_rms = hrv_time.loc[0, 'HRV_RMSSD'] # Feature
# # HRV Frequency-domain Indices
# # TODO: get NaN values within windows (*)
# # HRV_ULF *
# # HRV_LF *
# # HRV_HF
# # HRV_VHF
# # HRV_LFHF - Ratio LF/HF *
# # HRV_LFn *
# # HRV_HFn
# hrv_freq = nk.hrv_frequency(w_ecg_rpeaks, sampling_rate=sampling_rate, ulf=(0.01, 0.04), lf=(0.04, 0.15), hf=(0.15, 0.4), vhf=(0.4, 1.))
# hrv_ULF = hrv_freq.loc[0, 'HRV_ULF'] # Feature
# hrv_LF = hrv_freq.loc[0, 'HRV_LF'] # Feature
# hrv_HF = hrv_freq.loc[0, 'HRV_HF'] # Feature
# hrv_VHF = hrv_freq.loc[0, 'HRV_VHF'] # Feature
# hrv_lf_hf_ratio = hrv_freq.loc[0, 'HRV_LFHF'] # Feature
# hrv_f_sum = np.nansum(np.hstack((hrv_ULF, hrv_LF, hrv_HF, hrv_VHF)))
# # TODO: rel_f
# # hrv_rel_f =
# hrv_LFn = hrv_freq.loc[0, 'HRV_LFn'] # Feature
# hrv_HFn = hrv_freq.loc[0, 'HRV_HFn'] # Feature
# ==
## ECG
w_ecg_cleaned = ecg_cleaned[i: window + i]
_, ecg_info = nk.ecg_peaks(w_ecg_cleaned, sampling_rate=sampling_rate)
w_ecg_rpeaks = ecg_info['ECG_R_Peaks']
ecg_nni = pyhrv.tools.nn_intervals(w_ecg_rpeaks)
# HR
rs_hr = pyhrv.time_domain.hr_parameters(ecg_nni)
hr_mean = rs_hr['hr_mean'] # Feature
hr_std = rs_hr['hr_std'] # Feature
# HRV-time
rs_hrv = pyhrv.time_domain.nni_parameters(ecg_nni)
hrv_mean = rs_hrv['nni_mean'] # Feature
hrv_std = pyhrv.time_domain.sdnn(ecg_nni)['sdnn'] # Feature
rs_nn50 = pyhrv.time_domain.nn50(ecg_nni)
hrv_NN50 = rs_nn50['nn50'] # Feature
hrv_pNN50 = rs_nn50['pnn50'] # Feature
hrv_time = nk.hrv_time(w_ecg_rpeaks, sampling_rate=sampling_rate, show=False)
hrv_TINN = hrv_time.loc[0, 'HRV_TINN'] # Feature
hrv_rms = pyhrv.time_domain.rmssd(ecg_nni)['rmssd'] # Feature
# HRV-freq
hrv_freq = pyhrv.frequency_domain.welch_psd(ecg_nni, fbands={'ulf': (0.01, 0.04), 'vlf': (0.04, 0.15), 'lf': (0.15, 0.4), 'hf': (0.4, 1)}, mode='dev')
# hrv_freq = hrv_freq.as_dict()
hrv_freq = hrv_freq[0]
hrv_ULF = hrv_freq['fft_abs'][0] # Feature
hrv_LF = hrv_freq['fft_abs'][1] # Feature
hrv_HF = hrv_freq['fft_abs'][2] # Feature
hrv_VHF = hrv_freq['fft_abs'][3] # Feature
hrv_lf_hf_ratio = hrv_freq['fft_ratio'] # Feature
hrv_f_sum = hrv_freq['fft_total'] # Feature
hrv_rel_ULF = hrv_freq['fft_rel'][0] # Feature
hrv_rel_LF = hrv_freq['fft_rel'][1] # Feature
hrv_rel_HF = hrv_freq['fft_rel'][2] # Feature
hrv_rel_VHF = hrv_freq['fft_rel'][3] # Feature
hrv_LFn = hrv_freq['fft_norm'][0] # Feature
hrv_HFn = hrv_freq['fft_norm'][1] # Feature
# EDA
w_eda_data = eda_cleaned[i: window + i]
w_eda_phasic_tonic = eda_phasic_tonic[i: window + i]
eda_mean = np.mean(w_eda_data) # Feature
eda_std = np.std(w_eda_data) # Feature
eda_min = np.amin(w_eda_data) # Feature
eda_max = np.amax(w_eda_data) # Feature
# dynamic range: https://en.wikipedia.org/wiki/Dynamic_range
eda_slope = get_slope(w_eda_data) # Feature
eda_drange = eda_max / eda_min # Feature
eda_scl_mean = np.mean(w_eda_phasic_tonic['EDA_Tonic']) # Feature
eda_scl_std = np.std(w_eda_phasic_tonic['EDA_Tonic']) # Feature
eda_scr_mean = np.mean(w_eda_phasic_tonic['EDA_Phasic']) # Feature
eda_scr_std = np.std(w_eda_phasic_tonic['EDA_Phasic']) # Feature
eda_corr_scl_t = nk.cor(w_eda_phasic_tonic['EDA_Tonic'], w_eda_phasic_tonic['t'], show=False) # Feature
eda_scr_no = eda_scr_peaks['SCR_Peaks'][i: window + i].sum() # Feature
# Sum amplitudes in SCR signal
ampl = scr_info['SCR_Amplitude'][i: window + i]
eda_ampl_sum = np.sum(ampl[~np.isnan(ampl)]) # Feature
# TODO:
# eda_t_sum =
scr_peaks, scr_properties = scisig.find_peaks(w_eda_phasic_tonic['EDA_Phasic'], height=0)
width_scr = scisig.peak_widths(w_eda_phasic_tonic['EDA_Phasic'], scr_peaks, rel_height=0)
ht_scr = scr_properties['peak_heights']
eda_scr_area = 0.5 * np.matmul(ht_scr, width_scr[1]) # Feature
# EMG
## 5sec
w_emg_data = emg_filtered_dc[i: window + i]
emg_mean = np.mean(w_emg_data) # Feature
emg_std = np.std(w_emg_data) # Feature
emg_min = np.amin(w_emg_data)
emg_max = np.amax(w_emg_data)
emg_drange = emg_max / emg_min # Feature
emg_absint = np.abs(np.trapz(w_emg_data)) # Feature
emg_median = np.median(w_emg_data) # Feature
emg_perc_10 = np.percentile(w_emg_data, 10) # Feature
emg_perc_90 = np.percentile(w_emg_data, 90) # Feature
emg_peak_freq, emg_mean_freq, emg_median_freq = get_freq_features(w_emg_data) # Features
# TODO: PSD -> energy in seven bands
# emg_psd =
## 60 sec
peaks, properties = scisig.find_peaks(emg_filtered[i: window + i], height=0)
emg_peak_no = peaks.shape[0]
emg_peak_amp_mean = np.mean(properties['peak_heights']) # Feature
emg_peak_amp_std = np.std(properties['peak_heights']) # Feature
emg_peak_amp_sum = np.sum(properties['peak_heights']) # Feature
emg_peak_amp_max = np.abs(np.amax(properties['peak_heights']))
# https://www.researchgate.net/post/How_Period_Normalization_and_Amplitude_normalization_are_performed_in_ECG_Signal
emg_peak_amp_norm_sum = np.sum(properties['peak_heights'] / emg_peak_amp_max) # Feature
# Resp
w_resp_data = resp_processed[i: window + i]
## Inhalation / Exhalation duration analysis
idx = np.nan
count = 0
duration = dict()
first = True
for j in w_resp_data[~w_resp_data['RSP_Phase'].isnull()]['RSP_Phase'].to_numpy():
if j != idx:
if first:
idx = int(j)
duration[1] = []
duration [0] = []
first = False
continue
# print('New value', j, count)
duration[idx].append(count)
idx = int(j)
count = 0
count += 1
resp_inhal_mean = np.mean(duration[1]) # Feature
resp_inhal_std = np.std(duration[1]) # Feature
resp_exhal_mean = np.mean(duration[0]) # Feature
resp_exhal_std = np.std(duration[0]) # Feature
resp_inhal_duration = w_resp_data['RSP_Phase'][w_resp_data['RSP_Phase'] == 1].count()
resp_exhal_duration = w_resp_data['RSP_Phase'][w_resp_data['RSP_Phase'] == 0].count()
resp_ie_ratio = resp_inhal_duration / resp_exhal_duration # Feature
resp_duration = resp_inhal_duration + resp_exhal_duration # Feature
resp_stretch = w_resp_data['RSP_Amplitude'].max() - w_resp_data['RSP_Amplitude'].min() # Feature
resp_breath_rate = len(duration[1]) # Feature
## Volume: area under the curve of the inspiration phase on a respiratory cycle
resp_peaks, resp_properties = scisig.find_peaks(w_resp_data['RSP_Clean'], height=0)
resp_width = scisig.peak_widths(w_resp_data['RSP_Clean'], resp_peaks, rel_height=0)
resp_ht = resp_properties['peak_heights']
resp_volume = 0.5 * np.matmul(resp_ht, resp_width[1]) # Feature
# Temp
w_temp_data = data['Temp'][condition][i: window + i].flatten()
temp_mean = np.mean(w_temp_data) # Feature
temp_std = np.std(w_temp_data) # Feature
temp_min = np.amin(w_temp_data) # Feature
temp_max = np.amax(w_temp_data) # Feature
temp_drange = temp_max / temp_min # Feature
temp_slope = get_slope(w_temp_data.ravel()) # Feature
# chest_df_5 = chest_df_5.append({
# 'ACC_x_mean': acc_x_mean, 'ACC_y_mean': acc_y_mean, 'ACC_z_mean': acc_z_mean, 'ACC_xzy_mean': xyz_mean,
# 'ACC_x_std': acc_x_std, 'ACC_y_std': acc_y_std, 'ACC_z_std': acc_z_std, 'ACC_xyz_std': xyz_std,
# 'ACC_x_absint': acc_x_absint, 'ACC_y_absint': acc_y_absint, 'ACC_z_absint': acc_z_absint, 'ACC_xyz_absint': xyz_absint,
# 'ACC_x_peak': acc_x_peak, 'ACC_y_peak': acc_y_peak, 'ACC_z_peak': acc_z_peak,
# 'EMG_mean': emg_mean, 'EMG_std': emg_std, 'EMG_drange': emg_drange, 'EMG_absint': emg_absint, 'EMG_median': emg_median, 'EMG_perc_10': emg_perc_10,
# 'EMG_perc_90': emg_perc_90, 'EMG_peak_freq': emg_peak_freq, 'EMG_mean_freq': emg_mean_freq, 'EMG_median_freq': emg_median_freq
# }, ignore_index=True)
chest_df = chest_df.append({
'ACC_x_mean': acc_x_mean, 'ACC_y_mean': acc_y_mean, 'ACC_z_mean': acc_z_mean, 'ACC_xzy_mean': xyz_mean,
'ACC_x_std': acc_x_std, 'ACC_y_std': acc_y_std, 'ACC_z_std': acc_z_std, 'ACC_xyz_std': xyz_std,
'ACC_x_absint': acc_x_absint, 'ACC_y_absint': acc_y_absint, 'ACC_z_absint': acc_z_absint, 'ACC_xyz_absint': xyz_absint,
'ACC_x_peak': acc_x_peak, 'ACC_y_peak': acc_y_peak, 'ACC_z_peak': acc_z_peak,
'ECG_hr_mean': hr_mean, 'ECG_hr_std': hr_std, 'ECG_hrv_NN50': hrv_NN50, 'ECG_hrv_pNN50': hrv_pNN50, 'ECG_hrv_TINN': hrv_TINN, 'ECG_hrv_RMS': hrv_rms,
'ECG_hrv_ULF': hrv_ULF, 'ECG_hrv_LF': hrv_LF, 'ECG_hrv_HF': hrv_HF, 'ECG_hrv_VHF': hrv_VHF, 'ECG_hrv_LFHF_ratio': hrv_lf_hf_ratio, 'ECG_hrv_f_sum': hrv_f_sum,
'ECG_hrv_rel_ULF': hrv_rel_ULF, 'ECG_hrv_rel_LF': hrv_rel_LF, 'ECG_hrv_rel_HF': hrv_rel_HF, 'ECG_hrv_rel_VHF': hrv_rel_VHF, 'ECG_hrv_LFn': hrv_LFn, 'ECG_hrv_HFn': hrv_HFn,
'EDA_mean': eda_mean, 'EDA_std': eda_std, 'EDA_mean': eda_mean, 'EDA_min': eda_min, 'EDA_max': eda_max, 'EDA_slope': eda_slope,
'EDA_drange': eda_drange, 'EDA_SCL_mean': eda_scl_mean, 'EDA_SCL_std': eda_scl_mean, 'EDA_SCR_mean': eda_scr_mean, 'EDA_SCR_std': eda_scr_std,
'EDA_corr_SCL_t': eda_corr_scl_t, 'EDA_SCR_no': eda_scr_no, 'EDA_ampl_sum': eda_ampl_sum, 'EDA_scr_area': eda_scr_area,
'EMG_mean': emg_mean, 'EMG_std': emg_std, 'EMG_drange': emg_drange, 'EMG_absint': emg_absint, 'EMG_median': emg_median, 'EMG_perc_10': emg_perc_10,
'EMG_perc_90': emg_perc_90, 'EMG_peak_freq': emg_peak_freq, 'EMG_mean_freq': emg_mean_freq, 'EMG_median_freq': emg_median_freq,
'EMG_peak_no': emg_peak_no, 'EMG_peak_amp_mean': emg_peak_amp_mean, 'EMG_peak_amp_std': emg_peak_amp_std, 'EMG_peak_amp_sum': emg_peak_amp_sum,
'EMG_peak_amp_norm_sum': emg_peak_amp_norm_sum,
'RESP_inhal_mean': resp_inhal_mean, 'RESP_inhal_std': resp_inhal_std, 'RESP_exhal_mean': resp_exhal_mean, 'RESP_exhal_std': resp_exhal_std,
'RESP_ie_ratio': resp_ie_ratio, 'RESP_duration': resp_duration, 'RESP_stretch': resp_stretch, 'RESP_breath_rate': resp_breath_rate, 'RESP_volume': resp_volume,
'TEMP_mean': temp_mean, 'TEMP_std': temp_std, 'TEMP_min': temp_min, 'TEMP_max': temp_max, 'TEMP_drange': temp_drange, 'TEMP_slope': temp_slope
}, ignore_index=True)
# index += 1
# if index % 10 == 0:
# break
print('Elapsed Process', condition.shape[0], str(timedelta(seconds=time.time() - init)))
return chest_df, chest_df_5
def corr_and_featurize_ecg(self, recording, sample_freq, r_peaks, s_peaks,
q_peaks, p_peaks, t_peaks):
"""
Automatically derives features from ECG-files (only .dat files for now)
Args:
R-peaks
P-peaks
T-peaks
features (numpy array of str): an array of ECG-filenames in directory
labels (numpy array): an array of labels/diagnosis
directory (str): path to the features
demographical_data (DataFrame): A DataFrame containing feature name, age and gender
Returns:
features_out (DataFrame): A DataFrame with features for all ECG-records
"""
def interval_calc_simple(first_peak, second_peak, sample_freq):
try:
mean_interval = round((second_peak - first_peak).mean(), 5)
except:
mean_interval = float("NaN")
try:
std_interval = round((second_peak - first_peak).std(), 5)
except:
std_interval = float("NaN")
return mean_interval, std_interval
feature_list = []
feature_name = []
if len(r_peaks) and len(q_peaks) and len(s_peaks) and len(
p_peaks) and len(t_peaks) < 3:
try:
temp_data = nk.ecg_process(recording, sample_freq)[0]
r_peaks = np.where(temp_data['ECG_R_Peaks'] == 1)[0]
p_peaks = np.where(temp_data['ECG_P_Peaks'] == 1)[0]
q_peaks = np.where(temp_data['ECG_Q_Peaks'] == 1)[0]
s_peaks = np.where(temp_data['ECG_S_Peaks'] == 1)[0]
t_peaks = np.where(temp_data['ECG_T_Peaks'] == 1)[0]
p_onset = np.where(temp_data['ECG_P_Onsets'] == 1)[0]
t_offset = np.where(temp_data['ECG_T_Offsets'] == 1)[0]
clean_rec = temp_data['ECG_Clean']
analysis = True
except:
analysis = False
r_peaks = np.array([1, 2])
p_peaks = np.array([1, 2])
q_peaks = np.array([1, 2])
s_peaks = np.array([1, 2])
t_peaks = np.array([1, 2])
else:
analysis = True
clean_rec = nk.ecg_clean(recording)
try:
r_peaks = processing.peaks.correct_peaks(clean_rec,
r_peaks,
search_radius=25,
smooth_window_size=7,
peak_dir='compare')
except:
r_peaks = r_peaks
try:
q_peaks = processing.peaks.correct_peaks(clean_rec,
q_peaks,
search_radius=25,
smooth_window_size=7,
peak_dir='compare')
except:
q_peaks = q_peaks
try:
s_peaks = processing.peaks.correct_peaks(clean_rec,
s_peaks,
search_radius=25,
smooth_window_size=7,
peak_dir='compare')
except:
s_peaks = s_peaks
try:
t_peaks = processing.peaks.correct_peaks(clean_rec,
t_peaks,
search_radius=25,
smooth_window_size=7,
peak_dir='compare')
except:
t_peaks = t_peaks
try:
p_peaks = processing.peaks.correct_peaks(clean_rec,
p_peaks,
search_radius=25,
smooth_window_size=7,
peak_dir='compare')
except:
p_peaks = p_peaks
if self.rpeak_int == True:
feature_name.append("mean_rr_interval")
feature_name.append("sd_rr_interval")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append((np.diff(r_peaks) / sample_freq).mean())
feature_list.append((np.diff(r_peaks) / sample_freq).std())
if self.rpeak_amp == True:
feature_name.append("mean_r_peak")
feature_name.append("sd_r_peak")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append(recording[r_peaks].mean())
feature_list.append(recording[r_peaks].std())
if self.ppeak_int == True:
feature_name.append("mean_pp_interval")
feature_name.append("sd_pp_interval")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append((np.diff(p_peaks) / sample_freq).mean())
feature_list.append((np.diff(p_peaks) / sample_freq).std())
if self.ppeak_amp == True:
feature_name.append("mean_p_peak")
feature_name.append("sd_p_peak")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append(clean_rec[p_peaks].mean())
feature_list.append(clean_rec[p_peaks].std())
if self.tpeak_int == True:
feature_name.append("mean_tt_interval")
feature_name.append("sd_tt_interval")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append((np.diff(t_peaks) / sample_freq).mean())
feature_list.append((np.diff(t_peaks) / sample_freq).std())
if self.tpeak_amp == True:
feature_name.append("mean_t_peak")
feature_name.append("sd_t_peak")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append(clean_rec[t_peaks].mean())
feature_list.append(clean_rec[t_peaks].std())
if self.qpeak_int == True:
feature_name.append("mean_qq_interval")
feature_name.append("sd_qq_interval")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append((np.diff(q_peaks) / sample_freq).mean())
feature_list.append((np.diff(q_peaks) / sample_freq).std())
if self.qpeak_amp == True:
feature_name.append("mean_q_peak")
feature_name.append("sd_q_peak")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append(clean_rec[q_peaks].mean())
feature_list.append(clean_rec[q_peaks].std())
if self.speak_int == True:
feature_name.append("mean_q_peak")
feature_name.append("sd_q_peak")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append((np.diff(s_peaks) / sample_freq).mean())
feature_list.append((np.diff(s_peaks) / sample_freq).std())
if self.speak_amp == True:
feature_name.append("mean_s_peak")
feature_name.append("sd_s_peak")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
feature_list.append(clean_rec[s_peaks].mean())
feature_list.append(clean_rec[s_peaks].std())
if self.qrs_duration == True:
feature_name.append("qrs_mean")
feature_name.append("qrs_std")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
qrs_mean, qrs_std = interval_calc_simple(
q_peaks, s_peaks, sample_freq)
feature_list.append(qrs_mean)
feature_list.append(qrs_std)
if self.qt_duration == True:
feature_name.append("qt_mean")
feature_name.append("qt_std")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
qt_mean, qt_std = interval_calc_simple(q_peaks, t_peaks,
sample_freq)
feature_list.append(qt_mean)
feature_list.append(qt_std)
if self.pr_duration == True:
feature_name.append("pr_mean")
feature_name.append("pr_std")
if analysis == False:
feature_list.append(float("nan"))
feature_list.append(float("nan"))
elif analysis == True:
pr_mean, pr_std = interval_calc_simple(p_peaks, r_peaks,
sample_freq)
feature_list.append(pr_mean)
feature_list.append(pr_std)
feature_list = np.asarray(feature_list)
feature_name = np.asarray(feature_name)
return feature_list, feature_name, [
p_peaks, q_peaks, r_peaks, s_peaks, t_peaks
]
def get_12ECG_features_labels(data, header_data):
tmp_hea = header_data[0].split(' ')
ptID = tmp_hea[0]
num_leads = int(tmp_hea[1])
sample_Fs= int(tmp_hea[2])
gain_lead = np.zeros(num_leads)
for ii in range(num_leads):
tmp_hea = header_data[ii+1].split(' ')
gain_lead[ii] = int(tmp_hea[2].split('/')[0])
# for testing, we included the mean age of 57 if the age is a NaN
# This value will change as more data is being released
for iline in header_data:
if iline.startswith('#Age'):
tmp_age = iline.split(': ')[1].strip()
age = int(tmp_age if tmp_age != 'NaN' else 57)
elif iline.startswith('#Sex'):
tmp_sex = iline.split(': ')[1]
if tmp_sex.strip()=='Female':
sex =1
else:
sex=0
elif iline.startswith('#Dx'):
label = iline.split(': ')[1].split(',')[0]
signal = data[1]
gain = gain_lead[1]
N = len(signal)
sp= sample_Fs/N # resolución espectral
Y = np.fft.fft(signal*gain)
ff = np.linspace(0, (N/2)*sp, N/2).flatten()
fmax = float(ff[np.where(np.abs(Y[0:N//2]) == max(np.abs(Y[0:N//2])))])
# We are only using data from lead1
peaks,idx = detect_peaks(signal,sample_Fs,gain)
# mean
mean_RR = np.mean(idx/sample_Fs*1000)
mean_R_Peaks = np.mean(peaks*gain)
# median
median_RR = np.median(idx/sample_Fs*1000)
median_R_Peaks = np.median(peaks*gain)
# standard deviation
std_RR = np.std(idx/sample_Fs*1000)
std_R_Peaks = np.std(peaks*gain)
# variance
var_RR = stats.tvar(idx/sample_Fs*1000)
var_R_Peaks = stats.tvar(peaks*gain)
# Skewness
skew_RR = stats.skew(idx/sample_Fs*1000)
skew_R_Peaks = stats.skew(peaks*gain)
# Kurtosis
kurt_RR = stats.kurtosis(idx/sample_Fs*1000)
kurt_R_Peaks = stats.kurtosis(peaks*gain)
# RMSSD (HRV)
rmssd = np.sqrt(np.mean(np.square(np.diff(idx))))
# All Peaks
ecg_signal = nk.ecg_clean(signal*gain, sampling_rate=sample_Fs, method="biosppy")
_ , rpeaks = nk.ecg_peaks(ecg_signal, sampling_rate=sample_Fs)
try:
signal_peak, waves_peak = nk.ecg_delineate(ecg_signal, rpeaks, sampling_rate=sample_Fs)
t_peaks = waves_peak['ECG_T_Peaks']
p_peaks = waves_peak['ECG_P_Peaks']
q_peaks = waves_peak['ECG_Q_Peaks']
s_peaks = waves_peak['ECG_S_Peaks']
p_onsets = waves_peak['ECG_P_Onsets']
t_offsets = waves_peak['ECG_T_Offsets']
except ValueError:
print('Exception raised!')
pass
# T Peaks
t_peaks = np.asarray(t_peaks, dtype=float)
t_peaks = t_peaks[~np.isnan(t_peaks)]
t_peaks = [int(a) for a in t_peaks]
mean_T_Peaks = np.mean([signal[w] for w in t_peaks])
# P peaks
p_peaks = np.asarray(p_peaks, dtype=float)
p_peaks = p_peaks[~np.isnan(p_peaks)]
p_peaks = [int(a) for a in p_peaks]
mean_P_Peaks = np.mean([signal[w] for w in p_peaks])
# Q peaks
q_peaks = np.asarray(q_peaks, dtype=float)
q_peaks = q_peaks[~np.isnan(q_peaks)]
q_peaks = [int(a) for a in q_peaks]
mean_Q_Peaks = np.mean([signal[w] for w in q_peaks])
# S peaks
s_peaks = np.asarray(s_peaks, dtype=float)
s_peaks = s_peaks[~np.isnan(s_peaks)]
s_peaks = [int(a) for a in s_peaks]
mean_S_Peaks = np.mean([signal[w] for w in s_peaks])
# P Onsets
p_onsets = np.asarray(p_onsets, dtype=float)
# p_onsets = p_onsets[~np.isnan(p_onsets)]
mean_P_Onsets = np.mean(p_onsets/sample_Fs*1000)
# T Onsets
t_offsets = np.asarray(t_offsets, dtype=float)
# t_offsets = t_offsets[~np.isnan(t_offsets)]
mean_T_offsets = np.mean(t_offsets/sample_Fs*1000)
features = [age,sex,fmax,mean_RR,mean_R_Peaks,mean_T_Peaks,mean_P_Peaks,mean_Q_Peaks,mean_S_Peaks,median_RR,median_R_Peaks,std_RR,std_R_Peaks,var_RR,var_R_Peaks,skew_RR,skew_R_Peaks,kurt_RR,kurt_R_Peaks,mean_P_Onsets,mean_T_offsets,rmssd,label]
return features
def get_HRVs_values(data, header_data):
filter_lowcut = 0.001
filter_highcut = 15.0
filter_order = 1
tmp_hea = header_data[0].split(' ')
ptID = tmp_hea[0]
num_leads = int(tmp_hea[1])
sample_Fs= int(tmp_hea[2])
gain_lead = np.zeros(num_leads)
for ii in range(num_leads):
tmp_hea = header_data[ii+1].split(' ')
gain_lead[ii] = int(tmp_hea[2].split('/')[0])
# for testing, we included the mean age of 57 if the age is a NaN
# This value will change as more data is being released
for iline in header_data:
if iline.startswith('#Age'):
tmp_age = iline.split(': ')[1].strip()
age = int(tmp_age if tmp_age != 'NaN' else 57)
# age = int(tmp_age)
elif iline.startswith('#Sex'):
tmp_sex = iline.split(': ')[1]
if tmp_sex.strip()=='Female':
sex =1
else:
sex=0
elif iline.startswith('#Dx'):
label = iline.split(': ')[1].split(',')[0]
signal = data[1]
gain = gain_lead[1]
ecg_signal = nk.ecg_clean(signal*gain, sampling_rate=sample_Fs, method="biosppy")
_ , rpeaks = nk.ecg_peaks(ecg_signal, sampling_rate=sample_Fs)
hrv_time = nk.hrv_time(rpeaks, sampling_rate=sample_Fs)
peaks, idx = detect_peaks(signal, sample_Fs, gain)
# print(len(signal), len(idx))
rr_intervals = idx / (sample_Fs * 1000)
rr_intervals = pd.Series(rr_intervals)
rr_ma = rr_intervals.rolling(3)
try:
signal_peak, waves_peak = nk.ecg_delineate(ecg_signal, rpeaks, sampling_rate=sample_Fs)
p_peaks = waves_peak['ECG_P_Peaks']
except ValueError:
print('Exception raised!')
pass
p_peaks = np.asarray(p_peaks, dtype=float)
p_peaks = p_peaks[~np.isnan(p_peaks)]
p_peaks = [int(a) for a in p_peaks]
p_time = [x/sample_Fs for x in p_peaks]
p_diff = np.diff(p_time)
# mean_P_Peaks = np.mean([signal[w] for w in p_peaks])
hrv_time['var_P_time'] = stats.tvar(p_diff)
hrv_time['var_P_peaks'] = stats.tvar(signal[np.array(p_peaks)])
hrv_time['age'] = age
hrv_time['label'] = label
return hrv_time