def extractTimeDomain(self, x):
try:
nni = self.extractRR(x)
nniParams = td.nni_parameters(nni=nni)
nniSD = td.sdnn(nni=nni)
nniDiff = td.nni_differences_parameters(nni=nni)
nniDiffRM = td.rmssd(nni=nni)
nniDiffSD = td.sdsd(nni=nni)
hrParams = td.hr_parameters(nni=nni)
return np.array([nniParams["nni_mean"], nniParams["nni_counter"], nniSD["sdnn"],
nniDiff["nni_diff_mean"], nniDiffRM["rmssd"], nniDiffSD["sdsd"],
hrParams["hr_mean"], hrParams["hr_std"]])
except:
return np.array([])
def compute_features(nni):
features = {}
features['mean_hr'] = tools.heart_rate(nni).mean()
features['sdnn'] = td.sdnn(nni)[0]
features['rmssd'] = td.rmssd(nni)[0]
features['sdsd'] = td.sdsd(nni)[0]
features['nn20'] = td.nn20(nni)[0]
features['pnn20'] = td.nn20(nni)[1]
features['nn50'] = td.nn50(nni)[0]
features['pnn50'] = td.nn50(nni)[1]
features['hf_lf_ratio'] = fd.welch_psd(nni, show=False)['fft_ratio']
features['very_lf'] = fd.welch_psd(nni, show=False)['fft_peak'][0]
features['lf'] = fd.welch_psd(nni, show=False)['fft_peak'][1]
features['hf'] = fd.welch_psd(nni, show=False)['fft_peak'][2]
features['log_very_lf'] = fd.welch_psd(nni, show=False)['fft_log'][0]
features['log_lf'] = fd.welch_psd(nni, show=False)['fft_log'][1]
features['log_hf'] = fd.welch_psd(nni, show=False)['fft_log'][2]
features['sampen'] = nl.sample_entropy(nni)[0]
return features
def cal_hrv(data_list):
if (np.nan in data_list) or (len(data_list) <= 1):
sdsd = np.nan
rmssd = np.nan
sd1 = np.nan
sd2 = np.nan
sd_ratio = np.nan
a1 = np.nan
a2 = np.nan
a_ratio = np.nan
sampen = np.nan
else:
# sdsd
sdsd = round(td.sdsd(nni=data_list)['sdsd'], 5)
# rmssd
rmssd = round(td.rmssd(nni=data_list)['rmssd'], 5)
# sd1, sd2
nl_results = nl.poincare(nni=data_list)
sd1 = round(nl_results['sd1'], 5)
sd2 = round(nl_results['sd2'], 5)
sd_ratio = round(sd2 / sd1, 5)
# dfa a1, a2
dfa_results = nl.dfa(data_list)
try:
a1 = round(dfa_results['dfa_alpha1'], 5)
a2 = round(dfa_results['dfa_alpha2'], 5)
a_ratio = round(a2 / a1, 5)
except:
a1 = np.nan
print(a1, type(a1))
a2 = np.nan
print(a2, type(a2))
a_ratio = np.nan
# Sampen
t = np.std(data_list)
sampen = round(
nl.sample_entropy(nni=data_list, tolerance=t)['sampen'], 6)
return sdsd, rmssd, sd1, sd2, sd_ratio, a1, a2, a_ratio, sampen
def extractTimeDomain(self, x):
try:
nni = self.extractRR(x)
nniParams = td.nni_parameters(nni=nni)
nniSD = td.sdnn(nni=nni)
nniDiff = td.nni_differences_parameters(nni=nni)
nniDiffRM = td.rmssd(nni=nni)
nniDiffSD = td.sdsd(nni=nni)
hrParams = td.hr_parameters(nni=nni)
nn20 = td.nn20(nni=nni)
nn30 = td.nnXX(nni=nni, threshold=30)
nn50 = td.nn50(nni=nni)
# return np.array([nniParams["nni_mean"], nniParams["nni_counter"], nniSD["sdnn"],
# nniDiff["nni_diff_mean"], nniDiffRM["rmssd"], nniDiffSD["sdsd"],
# hrParams["hr_mean"], hrParams["hr_std"]])
return np.array([nniParams["nni_mean"], nniParams["nni_counter"], nniSD["sdnn"],
nniDiff["nni_diff_mean"], nniDiffRM["rmssd"], nniDiffSD["sdsd"],
hrParams["hr_mean"], hrParams["hr_std"], hrParams["hr_max"] - hrParams["hr_min"],
nn20["nn20"], nn20["pnn20"], nn30["nn30"], nn30["pnn30"], nn50["nn50"], nn50["pnn50"]])
except:
return np.array([])
def process(X):
features = []
plot = False
print(y[0:40])
tpls0 = []
tpls1 = []
tpls2 = []
tpls3 = []
shp = None
for i in range(len(X)):
_sample = X[i]
print(f"sample {i}: Class {y[i]}")
sample = _sample[~np.isnan(_sample)]
res = ecg(signal=sample, sampling_rate=300, show=False)
# FT
# N = len(sample) / 2
# T = 1.0 / 300.0
# xf = np.linspace(0.0, 1.0 / (2 * T), int(N // 2))
# yf = fft(res["filtered"])
# plt.plot(xf, 2.0 / N * np.abs(yf[0 : int(N // 2)]))
# plt.show()
median = calc_median(res["templates"])
# if (np.argmin(median) < 60) and not 0.7*np.max(median) > abs(np.min(median)):
if (
not np.max(median[55:65]) == np.max(median)
or (np.max(median) < -0.8 * np.min(median))
or (
not 0.75 * np.max(median) > -np.min(median)
and (
np.argmin(median) < 60
and (
np.min(median[:60]) < 1.5 * min(median[60:])
or np.min(median[60:]) < 1.5 * min(median[:60])
)
)
)
):
# and ((np.min(median) < 1.2 * np.min(
# median[[i for i in range(len(median)) if i != np.argmin(median)]])) or np.max(median[45:48]) > -np.min(median[65:75])):
# if np.min(median[45:55]) < np.min(median[0:45]) and np.min(median[65:80]) < np.min(median[80:]) and np.max(median[55:65]) == np.max(median):
# if np.max(median) < abs(np.min(median)) and np.min(median[50:55]) < np.min(median[60:65]):
# if abs(np.mean(median)) > abs(np.median(median)):
res = ecg(-sample, sampling_rate=300, show=False)
median = calc_median(res["templates"])
# neg = True
# res["templates"][j] = (res["templates"][j]-mean)/std
median = (median) / median.std()
if i < 40 and plot:
# plt.plot(res["templates"][j])
plt.title(y[i])
plt.plot(median)
plt.show()
# if not neg:
if y[i] == 0:
tpls0.append(median)
if y[i] == 1:
tpls1.append(median)
if y[i] == 2:
tpls2.append(median)
if y[i] == 3:
tpls3.append(median)
# beat characterization
heart_rate = res["heart_rate"]
filtered = res["filtered"]
rpeaks = res["rpeaks"]
# peaks in seconds, required by pyhrv
rpeaks_s = res["ts"][rpeaks]
qpeaks = np.array([find_minimum(filtered, r) for r in rpeaks])
speaks = np.array(
[find_minimum(filtered, r, direction="right") for r in rpeaks]
)
r_amplitude = filtered[rpeaks]
q_amplitude = filtered[qpeaks]
s_amplitude = filtered[speaks]
qrs_duration = speaks - qpeaks
# hrv_res = hrv(
# rpeaks=rpeaks_s,
# plot_tachogram=False,
# kwargs_ar={"order": 8},
# show=False,
# )
nni = time_domain.nni_parameters(rpeaks=rpeaks_s)
nni_diff = time_domain.nni_differences_parameters(rpeaks=rpeaks_s)
sdnn = time_domain.sdnn(rpeaks=rpeaks_s)
sdsd = time_domain.sdsd(rpeaks=rpeaks_s)
tri_index = time_domain.triangular_index(rpeaks=rpeaks_s, plot=False)
welch_psd = frequency_domain.welch_psd(rpeaks=rpeaks_s, mode="dev")[0]
# print(templates.shape, median.shape)
features.append(
build_features(q_amplitude, r_amplitude, s_amplitude, qrs_duration)
+ [
nni["nni_mean"],
nni["nni_min"],
nni["nni_max"],
nni_diff["nni_diff_mean"],
nni_diff["nni_diff_min"],
nni_diff["nni_diff_max"],
sdnn["sdnn"],
sdsd["sdsd"],
tri_index["tri_index"],
welch_psd["fft_ratio"],
]
+ list(welch_psd["fft_peak"] + welch_psd["fft_abs"] + welch_psd["fft_norm"])
)
# print(templates.shape, median.shape)
features = np.array(features)
print(f"computed features {features.shape}")
return features
def _computeSignal(self, signal):
obj = {}
# Best min_dist & thres for sphygmogram signal
peaks = peak.indexes(signal, min_dist=56, thres=0.16)
# Ignore un normal signls (with no peaks)
if (len(peaks) == 0): return obj
nn = tools.nn_intervals(peaks)
# Ignore un normal signls (with no NN)
if (len(nn) == 0): return
# Standard
obj = dict(td.nni_parameters(nn, peaks), **obj)
obj = dict(td.nni_differences_parameters(nn, peaks), **obj)
obj = dict(td.sdnn(nn, peaks), **obj)
obj = dict(td.sdnn_index(nn, peaks), **obj)
obj = dict(td.sdann(nn, peaks), **obj)
obj = dict(td.rmssd(nn, peaks), **obj)
obj = dict(td.sdsd(nn, peaks), **obj)
obj = dict(td.nn50(nn, peaks), **obj)
obj = dict(td.nn20(nn, peaks), **obj)
obj = dict(td.geometrical_parameters(nn, peaks, plot=False), **obj)
del obj['nni_histogram']
# Additional
obj = dict({'cv': self._cv(obj['sdnn'], obj['nni_mean'])}, **obj)
peaks_diff = tools.nni_diff(peaks)
obj = dict({'MxDMn': max(peaks_diff) - min(peaks_diff)}, **obj)
obj = dict({'MxRMn': max(peaks_diff) / min(peaks_diff)}, **obj)
obj = dict({'Mo': stats.mode(peaks_diff)[0][0]}, **obj)
counter = Counter(peaks_diff)
idx = list(counter.keys()).index(obj["Mo"])
obj = dict({'AMo': list(counter.values())[idx]}, **obj)
obj = dict({'SI': obj['AMo'] / (2 * obj['Mo'] * obj['MxDMn'])}, **obj)
# Autocorrelation function
# Frequency stats
welch = frequency_domain(signal).stats['welch']['params']
bands = list(welch['fft_bands'].keys())
obj = dict({'TP': welch['fft_total']}, **obj)
obj = dict({'HF': welch['fft_rel'][bands.index('hf')]}, **obj)
obj = dict({'LF': welch['fft_rel'][bands.index('lf')]}, **obj)
obj = dict({'VLF': welch['fft_rel'][bands.index('vlf')]}, **obj)
obj = dict({'ULF': welch['fft_rel'][bands.index('ulf')]}, **obj)
obj = dict({'HFav': welch['fft_abs'][bands.index('hf')]}, **obj)
obj = dict({'LFav': welch['fft_abs'][bands.index('lf')]}, **obj)
obj = dict({'VLFav': welch['fft_abs'][bands.index('vlf')]}, **obj)
obj = dict({'ULFav': welch['fft_abs'][bands.index('ulf')]}, **obj)
obj = dict({'(LF/HF)av': obj['LFav'] / obj['HFav']}, **obj)
obj = dict({'IC': obj['LF'] / obj['VLF']}, **obj)
for k in obj:
if (math.isnan(obj[k])):
obj[k] = 0
return obj
