def test_eda_peaks():
sampling_rate = 1000
eda = nk.eda_simulate(duration=30, sampling_rate=sampling_rate,
scr_number=6, noise=0, drift=0.01, random_state=42)
eda_phasic = nk.eda_phasic(nk.standardize(eda), method='highpass')["EDA_Phasic"].values
signals, info = nk.eda_peaks(eda_phasic, method="gamboa2008")
onsets, peaks, amplitudes = biosppy.eda.basic_scr(eda_phasic, sampling_rate=1000)
assert np.allclose((info["SCR_Peaks"] - peaks).mean(), 0, atol=1e-5)
signals, info = nk.eda_peaks(eda_phasic, method="kim2004")
onsets, peaks, amplitudes = biosppy.eda.kbk_scr(eda_phasic, sampling_rate=1000)
assert np.allclose((info["SCR_Peaks"] - peaks).mean(), 0, atol=1)
def extract_eda_features(eda: List[float], sampling_rate) -> pd.DataFrame:
# eda = MinMaxScaler().fit_transform(np.array(eda).reshape(-1, 1)).ravel()
HIGHCUT_FREQUENCY = 5 # defaults as BioSPPy
nyquist_freq = 2 * HIGHCUT_FREQUENCY / sampling_rate # Normalize frequency to Nyquist Frequency (Fs/2)
if 0 < nyquist_freq < 1:
eda = nk.eda_clean(eda, sampling_rate=sampling_rate, method='biosppy')
eda_decomposed = nk.eda_phasic(eda, sampling_rate=sampling_rate)
scr_peaks, info = nk.eda_peaks(eda_decomposed['EDA_Phasic'], sampling_rate=sampling_rate)
signals = pd.DataFrame({"EDA_Raw": eda})
signals = pd.concat([signals, eda_decomposed, scr_peaks], axis=1)
return signals
def standardized_eda_process(eda_signal, sampling_rate, method="neurokit"):
eda_cleaned = nk.eda_clean(eda_signal,
sampling_rate=sampling_rate,
method=method)
eda_cleaned_standardized = nk.standardize(
eda_cleaned) # only change to eda_process
eda_decomposed = nk.eda_phasic(eda_cleaned_standardized,
sampling_rate=sampling_rate)
peak_signal, info = nk.eda_peaks(eda_decomposed["EDA_Phasic"].values,
sampling_rate=sampling_rate,
method=method,
amplitude_min=0.1)
signals = pd.DataFrame({
"EDA_Raw": eda_signal,
"EDA_Clean": eda_cleaned,
"EDA_Standardized": eda_cleaned_standardized
})
signals = pd.concat([signals, eda_decomposed, peak_signal], axis=1)
return signals, info
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 extract_physio_features(physio):
"""
Actual workhorse function for extracting physiological features
:param physio: pandas dataframe
:return: pandas dataframe of physiological features
"""
# ecg
if physio_sanity(physio):
correct_rrs = physio.rr.values * 1000
# get basic statistics
warnings.filterwarnings("ignore", category=RuntimeWarning)
hr = calculate_statistics(physio.hr.values, 'hr')
rr = calculate_statistics(correct_rrs, 'rr')
gsr = calculate_statistics(physio.gsr.values, 'gsr')
temp = calculate_statistics(physio.temperature.values, 'temp')
# get hrv variables
warnings.filterwarnings("ignore", category=UserWarning)
hrv = get_all_hrv_features(correct_rrs)
hrv = hrv.drop(['mean_hr', 'max_hr', 'min_hr', 'std_hr', 'tinn'],
axis=1)
hrv.columns = ['hrv__' + c for c in hrv.columns]
warnings.filterwarnings("default", category=UserWarning)
# get eda features
eda = physio.gsr
eda = eda.rolling(3).mean().bfill().values
eda_decomposed = nk.eda_phasic(eda, sampling_rate=1, method='cvxeda')
# warnings.filterwarnings("ignore", category=RuntimeWarning)
try:
peak_signal, info = nk.eda_peaks(
eda_decomposed["EDA_Phasic"].values,
sampling_rate=1,
method='neurokit',
amplitude_min=0.2)
except ValueError: # if there are no peaks found
sig_cols = [
'SCR_Onsets', 'SCR_Peaks', 'SCR_Height', 'SCR_Amplitude',
'SCR_RiseTime', 'SCR_Recovery', 'SCR_RecoveryTime'
]
peak_signal = pd.DataFrame(np.zeros(7 * len(eda)).reshape(
len(eda), 7),
columns=sig_cols)
info = {s: [np.nan] for s in sig_cols}
warnings.filterwarnings("default", category=RuntimeWarning)
# combine decomposed signals and peak detection results
signals = pd.DataFrame({
"EDA_Raw": physio.gsr.values,
'EDA_Clean': eda
})
signals = pd.concat([signals, eda_decomposed, peak_signal], axis=1)
# nk.eda_plot(signals, 1)
# get features for phasic and tonic component
phasic = eda_phasic_feature_extraction(signals)
tonic = eda_tonic_feature_extraction(signals.EDA_Tonic.values)
tonic = pd.DataFrame.from_dict(tonic, orient='index').T
eda = pd.concat((phasic, tonic), axis=1)
return pd.concat((hr, hrv, rr, gsr, eda, temp), axis=1)
else:
return None