chriotte
Reputation: 39
Spectral analysis on HRV data with LombScargle in Python
I'm working with RR peaks and want to derive the frequency domain measures for HRV to recreate the results from the native C package by Physionet (WFDB tools). Both signal processing and spectral analysis are new fields for me, but after a long week with trial an error I've hacked together some code based on the Astropy module after trying several other solutions.
from astropy.stats import LombScargle
import random
dy = 0.1 * random.randint(1,100)
t = drive01["time"].values
y = drive01["intervals"].values
frequency, power = LombScargle(t, y,dy).autopower(minimum_frequency=0.0,maximum_frequency=4)
plt.plot(frequency, power)
This creates a plot that looks quite similar to the plot from Physionets package.
Physionets HRV tools with the code get_hrv makes this plot
Then by calculating common frequency domain measures I get quite different results.
Pxx = np.nan_to_num(power)
Fxx = np.nan_to_num(frequency)
ulf = 0.003
vlf = 0.04
lf = 0.15
hf = 0.4
Fs = 15.5 # the sampling rate of the drive file
# find the indexes corresponding to the VLF, LF, and HF bands
vlf_freq_band = (Fxx >= ulf) & (Fxx <= vlf)
lf_freq_band = (Fxx >= vlf) & (Fxx <= lf)
hf_freq_band = (Fxx >= lf) & (Fxx <= hf)
tp_freq_band = (Fxx >= 0) & (Fxx <= hf)
# Calculate the area under the given frequency band
dy = 1.0 / Fs
VLF = np.trapz(y=abs(Pxx[vlf_freq_band]), x=None, dx=dy)
LF = np.trapz(y=abs(Pxx[lf_freq_band]), x=None, dx=dy)
HF = np.trapz(y=abs(Pxx[hf_freq_band]), x=None, dx=dy)
TP = np.trapz(y=abs(Pxx[tp_freq_band]), x=None, dx=dy)
LF_HF = float(LF) / HF
Python
'HF': 0.10918703853414605,
'LF': 0.050074418080717789,
'LF/HF': 0.45861137689028925,
'TP': 0.20150514290250854,
'VLF': 0.025953350304821571
From the Physionet package:
TOT PWR = 0.0185973
VLF PWR = 0.00372733
LF PWR = 0.00472635
HF PWR = 0.0101436
LF/HF = 0.465944
When comparing the results it looks like this:
Python Physionet
TP 0.201505143 0.0185973 Quite similar + decimal dif
HF 0.109187039 0.0101436 Quite similar + decimal dif
LF 0.050074418 0.00472635 Quite similar + decimal dif
VLF 0.02595335 0.00372733 Not similar
LF/HF 0.458611377 0.465944 Quite similar
The calculations in Python are based on the code from another Stackoverflow post but the fix he got from the respondent is based on a python module I'm not able to get working and he is not using the Lomb Periodgram. I'm very open for trying something else as well, as long as its working with uneven samples. the data I'm working with is the drivedb from Physionet and I've used the Physionet packages to make a text file with RR peaks and time which is read into a Pandas DataFrame. The textfile can be found here
Upvotes: 1
Views: 1060
Answers (1)
Xue
Reputation: 11
LombScargle based on the Astropy caculate power different with C package by Physionet (WFDB tools). I write lombscargle again in python and result the same with C package by Physionet (WFDB tools).
import numpy as np
import os
import math
import csv
from itertools import zip_longest
import time
DATA_PATH = '/home/quangpc/Desktop/Data/PhysionetData/mitdb/'
class FreqDomainClass:
@staticmethod
def power(freq, mag):
lo = [0, 0.0033, 0.04, 0.15]
hi = [0.0033, 0.04, 0.15, 0.4]
pr = np.zeros(4)
nbands = 4
for index in range(0, len(freq)):
pwr = np.power(mag[index], 2)
for n in range(0, nbands):
if (freq[index] >= lo[n]) and freq[index] <= hi[n]:
pr[n] += pwr
break
return pr
@staticmethod
def avevar(y):
var = 0.0
ep = 0.0
ave = np.mean(y)
for i in range(len(y)):
s = y[i] - ave
ep += s
var += s * s
var = (var - ep * ep / len(y)) / (len(y) - 1)
return var
def lomb(self, t, h, ofac, hifac):
period = max(t) - min(t)
z = h - np.mean(h)
f = np.arange(1 / (period * ofac), hifac * len(h) / (2 * period), 1 / (period * ofac))
f = f[:int(len(f) / 2) + 1]
f = np.reshape(f, (len(f), -1))
w = 2 * np.pi * f
lenght_t = len(t)
t = np.reshape(t, (lenght_t, -1))
t = np.transpose(t)
tau = np.arctan2(np.sum(np.sin(2 * w * t), axis=1), np.sum(np.cos(2 * w * t), axis=1)) / (2 * w)
tau = np.diag(tau)
tau = np.reshape(tau, (len(tau), -1))
tau = np.tile(tau, (1, lenght_t))
cos = np.cos(w * (t - tau))
sin = np.sin(w * (t - tau))
pc = np.power(np.sum(z * cos, axis=1), 2)
ps = np.power(np.sum(z * sin, axis=1), 2)
cs = pc / np.sum(np.power(cos, 2), axis=1)
ss = ps / np.sum(np.power(sin, 2), axis=1)
p = cs + ss
pwr = self.avevar(h)
nout = len(h)
p = p / (2 * pwr)
p = p / (nout / (2.0 * pwr))
return f, np.sqrt(p)
def lomb_for(self, t, h, ofac, hifac):
period = max(t) - min(t)
f = np.arange(1 / (period * ofac), hifac * len(h) / (2 * period), 1 / (period * ofac))
f = f[:int(len(f) / 2) + 1]
z = h - np.mean(h)
p = np.zeros(len(f))
for i in range(len(f)):
w = 2 * np.pi * f[i]
if w > 0:
twt = 2 * w * t
y = sum(np.sin(twt))
x = sum(np.cos(twt))
tau = math.atan2(y, x) / (2 * w)
wtmt = w * (t - tau)
cs = np.power(sum(np.multiply(z, np.cos(wtmt))), 2) / sum(np.power((np.cos(wtmt)), 2))
ss = np.power(sum(np.multiply(z, np.sin(wtmt))), 2) / sum(np.power((np.sin(wtmt)), 2))
p[i] = cs + ss
else:
p[i] = np.power(sum(np.multiply(z, t)), 1) / sum(np.power(t), 1)
pwr = self.avevar(h)
nout = len(h)
p = p / (2 * pwr)
p = p / (nout / (2.0 * pwr))
return f, np.sqrt(p)
def freq_domain(self, time, rr_intervals):
frequency, mag0 = self.lomb(time, rr_intervals, 4.0, 2.0)
frequency = np.round(frequency, 8)
mag0 = mag0 / 2.0
mag0 = np.round(mag0, 8)
result = self.power(frequency, mag0)
return result[0], result[1], result[2], result[3], result[0] + result[1] + result[2] + result[3], \
result[2] / result[3]
def time_domain(time, rr_intervals, ann):
sum_rr = 0.0
sum_rr2 = 0.0
rmssd = 0.0
totnn = 0
totnnn = 0
nrr = 1
totrr = 1
nnx = 0
nnn = 0
lastann = ann[0]
lastrr = int(rr_intervals[0])
lenght = 300
t = float(time[0])
end = t + lenght
i = 0
ratbuf = np.zeros(2400)
avbuf = np.zeros(2400)
sdbuf = np.zeros(2400)
for x in range(1, len(ann)):
t = float(time[x])
while t > (end+lenght):
i += 1
end += lenght
if t >= end:
if nnn > 1:
ratbuf[i] = nnn/nrr
sdbuf[i] = np.sqrt(((sdbuf[i] - avbuf[i]*avbuf[i]/nnn) / (nnn-1)))
avbuf[i] /= nnn
i += 1
nnn = nrr = 0
end += lenght
nrr += 1
totrr += 1
if ann[x] == 'N' and ann[x-1] == 'N':
rr_intervals[x] = int(rr_intervals[x])
nnn += 1
avbuf[i] += rr_intervals[x]
sdbuf[i] += (rr_intervals[x] * rr_intervals[x])
sum_rr += rr_intervals[x]
sum_rr2 += (rr_intervals[x] * rr_intervals[x])
totnn += 1
if lastann == 'N':
totnnn += 1
rmssd += (rr_intervals[x] - lastrr) * (rr_intervals[x] - lastrr)
# nndif[0] = NNDIF
if abs(rr_intervals[x] - lastrr) - 0.05 > (10 ** -10):
nnx += 1
lastann = ann[x-1]
lastrr = rr_intervals[x]
if totnn == 0:
return 0, 0, 0, 0
sdnn = np.sqrt((sum_rr2 - sum_rr * sum_rr / totnn) / (totnn - 1))
rmssd = np.sqrt(rmssd/totnnn)
pnn50 = nnx / totnnn
if nnn > 1:
ratbuf[i] = nnn / nrr
sdbuf[i] = np.sqrt((sdbuf[i] - avbuf[i] * avbuf[i] / nnn) / (nnn - 1))
avbuf[i] /= nnn
nb = i + 1
sum_rr = 0.0
sum_rr2 = 0.0
k = 0
h = 0
while k < nb:
if ratbuf[k] != 0:
h += 1
sum_rr += avbuf[k]
sum_rr2 += (avbuf[k] * avbuf[k])
k += 1
sdann = np.sqrt((sum_rr2 - sum_rr * sum_rr / h) / (h - 1))
return sdnn, sdann, rmssd, pnn50
def get_result_from_get_hrv(filename):
with open(filename, 'r') as f:
csv_reader = csv.reader(f, delimiter=',')
index = 0
for row in csv_reader:
if index > 0:
output = [s.strip() for s in row[0].split('=') if s]
# print('output = ', output)
if output[0] == 'SDNN':
sdnn = output[1]
if output[0] == 'SDANN':
sdann = output[1]
if output[0] == 'rMSSD':
rmssd = output[1]
if output[0] == 'pNN50':
pnn50 = output[1]
if output[0] == 'ULF PWR':
ulf = output[1]
if output[0] == 'VLF PWR':
vlf = output[1]
if output[0] == 'LF PWR':
lf = output[1]
if output[0] == 'HF PWR':
hf = output[1]
if output[0] == 'TOT PWR':
tp = output[1]
if output[0] == 'LF/HF':
ratio_lf_hf = output[1]
index += 1
return float(sdnn), float(sdann), float(rmssd), float(pnn50), float(ulf), float(vlf), \
float(lf), float(hf), float(tp), float(ratio_lf_hf)
def save_file():
extension = "atr"
result_all = []
file_process = ['File']
sdnn_l = ['sdnn']
sdann_l = ['sdann']
rmssd_l = ['rmssd']
pnn50_l = ['pnn50']
ulf_l = ['ulf']
vlf_l = ['vlf']
lf_l = ['lf']
hf_l = ['hf']
tp_l = ['tp']
ratio_lf_hf_l = ['ratio_lf_hf']
sdnn_l_p = ['sdnn']
sdann_l_p = ['sdann']
rmssd_l_p = ['rmssd']
pnn50_l_p = ['pnn50']
ulf_l_p = ['ulf']
vlf_l_p = ['vlf']
lf_l_p = ['lf']
hf_l_p = ['hf']
tp_l_p = ['tp']
ratio_lf_hf_l_p = ['ratio_lf_hf']
test_file = ['103', '113', '117', '121', '123', '200', '202', '210', '212', '213',
'219', '221', '213', '228', '231', '233', '234',
'101', '106', '108', '112', '114', '115', '116', '119', '122', '201', '203',
'205', '208', '209', '215', '220', '223', '230',
'105', '100']
file_dis = ['109', '111', '118', '124', '207', '214', '232']
for root, dirs, files in os.walk(DATA_PATH):
files = np.sort(files)
for name in files:
if extension in name:
if os.path.basename(name[:-4]) not in test_file:
continue
print('Processing file...', os.path.basename(name))
cur_dir = os.getcwd()
os.chdir(DATA_PATH)
os.system('rrlist {} {} -M -s >{}.rr'.format(extension, name.split('.')[0], name.split('.')[0]))
time_m = []
rr_intervals = []
ann = []
with open(name.split('.')[0] + '.rr', 'r') as rr_file:
for line in rr_file:
time_m.append(line.split(' ')[0])
rr_intervals.append(line.split(' ')[1])
ann.append(line.split(' ')[2].split('\n')[0])
time_m = np.asarray(time_m, dtype=float)
rr_intervals = np.asarray(rr_intervals, dtype=float)
sdnn, sdann, rmssd, pnn50 = time_domain(time_m, rr_intervals, ann)
if sdnn == 0 and sdann == 0 and rmssd == 0 and pnn50 == 0:
print('No result hrv')
file_dis.append(os.path.basename(name[:-4]))
continue
print('sdnn', sdnn)
print('rmssd', rmssd)
print('pnn50', pnn50)
print('sdann', sdann)
time_m = time_m - time_m[0]
time_m = np.round(time_m, 3)
time_nn = []
nn_intervals = []
for i in range(1, len(ann)):
if ann[i] == 'N' and ann[i - 1] == 'N':
nn_intervals.append(rr_intervals[i])
time_nn.append(time_m[i])
time_nn = np.asarray(time_nn, dtype=float)
nn_intervals = np.asarray(nn_intervals, dtype=float)
fc = FreqDomainClass()
ulf, vlf, lf, hf, tp, ratio_lf_hf = fc.freq_domain(time_nn, nn_intervals)
sdnn_l.append(sdnn)
sdann_l.append(sdann)
rmssd_l.append(rmssd)
pnn50_l.append(pnn50)
ulf_l.append(ulf)
vlf_l.append(vlf)
lf_l.append(lf)
hf_l.append(hf)
tp_l.append(tp)
ratio_lf_hf_l.append(ratio_lf_hf)
print('ULF PWR: ', ulf)
print('VLF PWR: ', vlf)
print('LF PWR: ', lf)
print('HF PWR: ', hf)
print('TOT PWR: ', tp)
print('LF/HF: ', ratio_lf_hf)
if os.path.exists('physionet_hrv.txt'):
os.remove('physionet_hrv.txt')
os.system('get_hrv -R ' + name.split('.')[0] + '.rr >> ' + 'physionet_hrv.txt')
sdnn, sdann, rmssd, pnn50, ulf, vlf, lf, hf, tp, ratio_lf_hf = \
get_result_from_get_hrv('physionet_hrv.txt')
file_process.append(os.path.basename(name[:-4]))
sdnn_l_p.append(sdnn)
sdann_l_p.append(sdann)
rmssd_l_p.append(rmssd)
pnn50_l_p.append(pnn50)
ulf_l_p.append(ulf)
vlf_l_p.append(vlf)
lf_l_p.append(lf)
hf_l_p.append(hf)
tp_l_p.append(tp)
ratio_lf_hf_l_p.append(ratio_lf_hf)
os.chdir(cur_dir)
result_all.append(file_process)
result_all.append(sdnn_l)
result_all.append(sdnn_l_p)
result_all.append(sdann_l)
result_all.append(sdann_l_p)
result_all.append(rmssd_l)
result_all.append(rmssd_l_p)
result_all.append(pnn50_l)
result_all.append(pnn50_l_p)
result_all.append(ulf_l)
result_all.append(ulf_l_p)
result_all.append(vlf_l)
result_all.append(vlf_l_p)
result_all.append(lf_l)
result_all.append(lf_l_p)
result_all.append(hf_l)
result_all.append(hf_l_p)
result_all.append(tp_l)
result_all.append(tp_l_p)
result_all.append(ratio_lf_hf_l)
result_all.append(ratio_lf_hf_l_p)
print(file_dis)
with open('hrv2.csv', 'w+') as f:
writer = csv.writer(f)
for values in zip_longest(*result_all):
writer.writerow(values)
def main():
extension = "atr"
for root, dirs, files in os.walk(DATA_PATH):
files = np.sort(files)
for name in files:
if extension in name:
if name not in ['101.atr']:
continue
cur_dir = os.getcwd()
os.chdir(DATA_PATH)
os.system('rrlist {} {} -M -s >{}.rr'.format(extension, name.split('.')[0], name.split('.')[0]))
time_m = []
rr_intervals = []
ann = []
with open(name.split('.')[0] + '.rr', 'r') as rr_file:
for line in rr_file:
time_m.append(line.split(' ')[0])
rr_intervals.append(line.split(' ')[1])
ann.append(line.split(' ')[2].split('\n')[0])
time_m = np.asarray(time_m, dtype=float)
rr_intervals = np.asarray(rr_intervals, dtype=float)
sdnn, sdann, rmssd, pnn50 = time_domain(time_m, rr_intervals, ann)
if sdnn == 0 and sdann == 0 and rmssd == 0 and pnn50 == 0:
print('No result hrv')
return 0
print('sdnn', sdnn)
print('rmssd', rmssd)
print('pnn50', pnn50)
print('sdann', sdann)
time_m = time_m - time_m[0]
time_m = np.round(time_m, 3)
time_nn = []
nn_intervals = []
for i in range(1, len(ann)):
if ann[i] == 'N' and ann[i - 1] == 'N':
nn_intervals.append(rr_intervals[i])
time_nn.append(time_m[i])
time_nn = np.asarray(time_nn, dtype=float)
nn_intervals = np.asarray(nn_intervals, dtype=float)
start = time.time()
fc = FreqDomainClass()
ulf, vlf, lf, hf, tp, ratio_lf_hf = fc.freq_domain(time_nn, nn_intervals)
end = time.time()
print('ULF PWR: ', ulf)
print('VLF PWR: ', vlf)
print('LF PWR: ', lf)
print('HF PWR: ', hf)
print('TOT PWR: ', tp)
print('LF/HF: ', ratio_lf_hf)
print('finish', end - start)
os.chdir(cur_dir)
Upvotes: 1
Related Questions
- my power spectra are believable? a comparison between lomb-scargle and fft (scipy.signal and numpy.fft)
- ECG Data Analysis on a real-time signal in Python
- signal.spectrogram returns too many hz
- scipy.signal.spectrogram frequency resolution
- Simple ECG waveform samples?
- fft of a biosignal of unknown frequencyin python
- Use of scipy.signal.lombscargle
- Filter design and frequency extraction in Python
- Correct way to use scipy.signal.spectral.lombscargle
- Using scipy.signal.spectral.lombscargle for period discovery