def run(inp, opt, cfg):
srate = inp['ecg']['srate']
data = arr.interp_undefined(inp['ecg']['vals'])
ret = arr.remove_wander_spline(data, srate)
return [{"srate": srate, "vals": ret}]
def run(inp, opt, cfg):
data = arr.interp_undefined(inp['ecg']['vals'])
srate = inp['ecg']['srate']
r_list = arr.detect_qrs(data, srate) # detect r-peak
ret_rpeak = []
for idx in r_list:
dt = idx / srate
ret_rpeak.append({'dt': dt, 'val': 1})
return [ret_rpeak]
def run(inp, opt, cfg):
data = arr.interp_undefined(inp['eeg']['vals'])
data -= smooth(np.array(data))
srate = int(inp['eeg']['srate'])
nfft = srate * 2 # srate * epoch size
fres = srate / nfft # frequency resolution (hz)
# frequency domain analysis
EPOCH_SIZE = int(srate * 2)
STRIDE_SIZE = int(srate * 0.5)
ps = []
for epoch_start in range(0, len(data) - EPOCH_SIZE + 1, STRIDE_SIZE): # 0.5초 마다 겹침
epoch_w = data[epoch_start:epoch_start + EPOCH_SIZE] # 2초 epoch
epoch_w = (epoch_w - np.mean(epoch_w)) * np.blackman(EPOCH_SIZE) # detrend and windowing
dft = np.fft.fft(epoch_w)[:srate] # 실수를 fft 했으므로 절반만 필요하다
dft[0] = 0 # dc 성분은 지움
ps.append(2 * np.abs(dft) ** 2) # 파워의 절대값인데 절반 날렸으므로
ps = np.mean(np.array(ps), axis=0)
pssum = np.cumsum(ps) # cummulative sum
pssum = pssum[1:]
totpow = pssum[fromhz(30, fres)]
sef = tohz(np.argmax(pssum > 0.95 * totpow), fres)
mf = tohz(np.argmax(pssum > 0.5 * totpow), fres)
delta = pssum[fromhz(4, fres) - 1] / pssum[-1] * 100
theta = (pssum[fromhz(8, fres) - 1] - pssum[fromhz(4, fres)]) / pssum[-1] * 100
alpha = (pssum[fromhz(12, fres) - 1] - pssum[fromhz(8, fres)]) / pssum[-1] * 100
beta = (pssum[fromhz(30, fres) - 1] - pssum[fromhz(12, fres)]) / pssum[-1] * 100
gamma = (pssum[-1] - pssum[fromhz(30, fres)]) / pssum[-1] * 100
# pttmax_list.append()
# pttdmax_list.append({'dt': dmax_dt, 'val': (dmax_dt - rpeak_dt) * 1000})
# pttmin_list.append({'dt': min_dt, 'val': (min_dt - rpeak_dt) * 1000})
#
return [
[{'dt': cfg['interval'], 'val': 10 * np.log10(totpow)}],
[{'dt': cfg['interval'], 'val': sef}],
[{'dt': cfg['interval'], 'val': mf}],
[{'dt': cfg['interval'], 'val': delta}],
[{'dt': cfg['interval'], 'val': theta}],
[{'dt': cfg['interval'], 'val': alpha}],
[{'dt': cfg['interval'], 'val': beta}],
[{'dt': cfg['interval'], 'val': gamma}]
]
def run(inp, opt, cfg):
"""
http:#ocw.utm.my/file.php/38/SEB4223/07_ECG_Analysis_1_-_QRS_Detection.ppt%20%5BCompatibility%20Mode%5D.pdf
"""
global hist_ppga, hist_hbi
data = arr.interp_undefined(inp['pleth']['vals'])
srate = inp['pleth']['srate']
minlist, maxlist = arr.detect_peaks(data, srate) # extract beats
beat_res = [{'dt': idx / srate, 'val': 1} for idx in maxlist]
ppga_res = []
hbi_res = []
ppga_perc_res = []
hbi_perc_res = []
spi_res = []
for i in range(len(maxlist) - 1):
dt = maxlist[i + 1] / srate
hbi = (maxlist[i + 1] - maxlist[i]) / srate * 1000
ppga = data[maxlist[i + 1]] - data[minlist[i]]
#hbi_perc = hist_hbi.percentile(hbi) * 0.7 + st.norm.cdf(hbi, 754.7, 210.8) * 30
hbi_perc = hist_hbi.percentile(hbi) * 0.7 + st.norm.cdf(hbi, 700,
100) * 30
#ppga_perc = hist_ppga.percentile(ppga) * 0.7 + st.norm.cdf(ppga, 2.428, 1.896) * 30
ppga_perc = hist_ppga.percentile(ppga) * 0.7 + st.norm.cdf(
ppga, 1, 0.2) * 30
# hbi_perc = hist_hbi.percentile(hbi) * 0.7 + hist_hbi_grp.percentile(hbi) * 0.3
# ppga_perc = hist_ppga.percentile(ppga) * 0.7 + hist_ppga_grp.percentile(ppga) * 0.3
spi = 100 - (0.7 * ppga_perc + 0.3 * hbi_perc)
ppga_res.append({'dt': dt, 'val': ppga})
hbi_res.append({'dt': dt, 'val': hbi})
ppga_perc_res.append({'dt': dt, 'val': ppga_perc})
hbi_perc_res.append({'dt': dt, 'val': hbi_perc})
spi_res.append({'dt': dt, 'val': spi})
hist_hbi.learn(hbi)
hist_ppga.learn(ppga)
return [beat_res, ppga_res, hbi_res, ppga_perc_res, hbi_perc_res, spi_res]
def run(inp, opt, cfg):
data = arr.interp_undefined(inp['ecg']['vals'])
srate = inp['ecg']['srate']
span = int(srate * 10) # Cut every 10 second regardless of interval
baseline = [0] * len(data)
x = np.arange(0, span)
for spos in range(0, len(data), span): # start position of this segment
if spos + span > len(data):
span = len(data) - spos
x = x[0:span]
med = np.median(
data[spos:spos +
span]) # compute overall median for the entire waveform
for j in range(span):
data[
spos +
j] -= med # shift each sample of the entire waveform by this median value
y = data[spos:spos + span]
p = np.polyfit(x, y, 4)
y = np.polyval(p, x)
for j in range(span):
baseline[spos + j] = y[j]
data[spos + j] -= y[j]
r_list = arr.detect_qrs(data, srate) # detect r-peak
for i in range(len(r_list) - 1): # for each rr interval
idx1 = r_list[i]
idx2 = r_list[i + 1]
if idx1 + 1 > idx2:
continue
med = np.median(data[idx1 + 1:idx2])
for j in range(idx1 + 1, idx2):
data[j] -= med
return [{'srate': srate, 'vals': data}, {'srate': srate, 'vals': baseline}]
def run(inp, opt, cfg):
"""
calculate ppv from arterial waveform
:param art: arterial waveform
:return: max, min, upper envelope, lower envelope, respiratory rate, ppv
"""
vsrate = inp['volume']['srate']
psrate = inp['awp']['srate']
fsrate = inp['flow']['srate']
if vsrate != psrate or vsrate != fsrate:
print("sampling rates of volume, flow and awp are different")
return
srate = vsrate
vdata = arr.interp_undefined(inp['volume']['vals'])
fdata = arr.interp_undefined(inp['flow']['vals'])
pdata = arr.interp_undefined(inp['awp']['vals'])
# if srate < 200:
# vdata = arr.resample_hz(vdata, srate, 200)
# fdata = arr.resample_hz(fdata, srate, 200)
# pdata = arr.resample_hz(pdata, srate, 200)
# srate = 200
vdata = np.array(vdata)
fdata = np.array(fdata) / 60 # L/min -> L/sec
pdata = np.array(pdata)
#fdata = np.diff(vdata) * srate / 1000 # make difference to rate
#vdata = vdata[:-1] # remove the last sample
#pdata = pdata[:-1] # remove the last sample
vmax = max(vdata)
vmin = min(vdata)
v95 = vmax - (vmax - vmin) * 0.1
v5 = vmin + (vmax - vmin) * 0.1
vret = []
cret = []
rret = []
p0ret = []
nstep = 31
vstep = (v95 - v5) / nstep
for i in range(nstep):
# collect data
vfrom = v5 + vstep * i
seg_idx = np.logical_and(vfrom < vdata, vdata <= vfrom + vstep)
if sum(seg_idx) < 3:
print('number of samples in data seg < 3')
continue
pseg = pdata[seg_idx]
vseg = vdata[seg_idx]
fseg = fdata[seg_idx]
A = np.vstack([vseg, fseg, np.ones(len(vseg))]).T
cinv, r, p0 = np.linalg.lstsq(A, pseg)[0]
c = 1 / cinv
vret.append({'dt': i * 0.02, 'val': vfrom})
cret.append({'dt': i * 0.02, 'val': c})
rret.append({'dt': i * 0.02, 'val': r})
p0ret.append({'dt': i * 0.02, 'val': p0})
return [
#{'dt':0, 'srate':srate, 'vals':list(fdata)},
vret,
cret,
rret,
p0ret
]
def run(inp, opt, cfg):
data = arr.interp_undefined(inp['ecg']['vals'])
srate = inp['ecg']['srate']
ecg_500 = data
if srate != 500:
ecg_500 = arr.resample(data, math.ceil(len(data) / srate *
500)) # resample to 500 Hz
srate = 500
ecg_filt = arr.band_pass(ecg_500, srate, 0.01, 100) # filtering
ecg_filt = arr.remove_wander_spline(ecg_filt,
srate) # remove baseline wander
r_list = arr.detect_qrs(ecg_filt, srate) # detect r-peak
new_r_list = []
for ridx in r_list: # remove qrs before and after overlap
if cfg['overlap'] <= ridx / srate:
new_r_list.append(ridx)
r_list = new_r_list
ret_rpeak = []
for ridx in r_list:
ret_rpeak.append({'dt': ridx / srate})
segbeats = 128
segsteps = 32 # int(segbeats/4)
# for each segments
twavs = []
twars = []
ret_twav = []
ret_twar = []
ret_avg_beat = {'srate': srate, 'vals': [0] * len(ecg_500)}
iseg = 0
for seg_start in range(
0,
len(r_list) - segbeats, segsteps
): # Separates in 128-beat units regardless of input length
iseg += 1
hrs = [] # calculate hrs
for i in range(segbeats - 1):
hr = srate / (r_list[seg_start + i + 1] - r_list[seg_start + i])
hrs.append(hr)
if max(hrs) - min(hrs) > 20:
# print('seg ' + iseg + ' excluded HR diff > ' + diff_hr)
continue
# only -250 to 350 ms from R peak
idx_r = int(0.25 * srate) # idx_r == 125
beat_len = int(0.6 * srate) # beat_len == 300
beats = []
for i in range(segbeats):
ridx = r_list[seg_start + i]
beat = ecg_filt[ridx - idx_r:ridx - idx_r + beat_len]
beats.append(beat)
beats = np.array(beats)
# remove each beat's baseline voltage
# no effect because of R peak leveling is below
# Baseline correction included estimation of the baseline in the isoelectric PQ
# segment by averaging 16 successive samples in this time window
pq_width = int(0.008 * srate)
# for i in range(segbeats):
# idx_base = arr.min_idx(beats[i], idx_r - int(0.15 * srate), idx_r)
# min_std = 999999
# for j in range(idx_base - int(0.03 * srate), idx_base + int(0.03 * srate)):
# # The baseline is the point at which the standard deviation of around 15ms is minimized.
# this_std = np.std(beats[i][j - pq_width:j + pq_width])
# if this_std < min_std:
# idx_base = j
# min_std = this_std
# beats[i] -= np.mean(beats[i][idx_base - pq_width:idx_base + pq_width])
# calculate average beat
avg_beat = np.mean(beats, axis=0) # average beat of the segbeats beats
# find minimum values from avg_beat in both sides
idx_start = idx_r - int(0.15 * srate) # idx_start == 50
idx_end = idx_r + int(0.1 * srate) # idx_end == 175
idx_base = arr.min_idx(avg_beat, idx_start,
idx_r) # avg_beat's baseline
min_std = 999999 # find minimum std value
for j in range(idx_base - int(0.03 * srate),
idx_base + int(0.03 * srate)):
idx_from = max(0, j - pq_width)
idx_to = min(len(avg_beat), j + pq_width)
this_std = np.std(avg_beat[idx_from:idx_to])
# print("{} {}".format(j, this_std))
if this_std < min_std:
idx_base = j
min_std = this_std
# print("idx_base={}", idx_base)
min_left = np.mean(avg_beat[idx_base - pq_width:idx_base + pq_width])
min_right = np.min(avg_beat[idx_r:idx_end])
# threshold = 5% of max val
th_left = min_left + 0.05 * (avg_beat[idx_r] - min_left)
th_right = min_right + 0.05 * (avg_beat[idx_r] - min_right)
idx_qrs_start = idx_r - int(0.05 * srate)
idx_qrs_end = idx_r + int(0.05 * srate)
for j in range(idx_r, idx_r - int(0.1 * srate), -1): # idx_r = 125
if avg_beat[j] < th_left:
idx_qrs_start = j
break
for j in range(idx_r, idx_r + int(0.1 * srate)):
if avg_beat[j] < th_right:
idx_qrs_end = j
break
# find offset with maximum correlation
offsets = [] # for each beat, likes [0, -1, 0, 0, 1, ...]
qrs_coeffs = []
offset_width = int(0.01 * srate) # 3 = range for finding offset
for i in range(segbeats): # for each beat
maxoffset = -offset_width
maxce = -999999
for offset in range(-offset_width, offset_width + 1):
ce = arr.corr(
avg_beat[idx_qrs_start:idx_qrs_end],
beats[i][offset + idx_qrs_start:offset + idx_qrs_end])
if maxce < ce:
maxoffset = offset
maxce = ce
offsets.append(maxoffset)
qrs_coeffs.append(maxce)
# move beats by the offset
new_beats = []
for i in range(segbeats):
ost = offsets[i]
beat = beats[i].tolist()
if ost < 0:
beat = [0] * -ost + beat[:ost]
else:
beat = beat[ost:] + [0] * ost
new_beats.append(beat)
beats = np.array(new_beats) # beats.shape == (segbeats,300)
# calculate average beat
avg_beat = np.mean(beats, axis=0) # average beat of the segbeats beats
# replace vpc as template
nreplaced = 0
for i in range(segbeats):
ce = arr.corr(avg_beat, beats[i])
if ce < 0.95:
nreplaced += 1
beats[i] = copy.deepcopy(avg_beat)
offsets[i] = 0
#print('{} beats are replaced'.format(nreplaced))
if nreplaced > 0.1 * segbeats:
print('excluded VPC > {}'.format(nreplaced))
continue
# qrs level alignment
# idx_r == 125
# len(avg_beat) == beat_len == 300
for i in range(segbeats):
beats[i] -= beats[i][idx_r]
# plot for debugging
# plt.plot()
# for i in range(segbeats):
# col = 'blue'
# if i % 2:
# col = 'red'
# plt.plot(beats[i], c=col, ls='-')
# plt.savefig('{:02d}_{}.png'.format(opt['ifile'], iseg))
# plt.close()
# gather segbeats beats from idx_r(125) to beat_len(300)
# power spectrums of segbeats beats
spect = []
for idx_from_r in range(beat_len - idx_r):
timed_samples = beats[:, idx_r + idx_from_r]
# timed_samples *= np.hamming(len(timed_samples))
segfft = 2**np.abs(np.fft.fft(timed_samples))
spect.append(segfft) # each segbeats beat fft result
spect = np.array(
spect) # rows == idx_from_r, cols == frequency(0-segbeats)
# power spectra are summed into a composite in which
# the magnitude at 0.5 cycles/beat indicates raw alternans (in mv2)
# cum_spect.shape == segbeats
# cumulative spectum of beats
st_start = int(0.1 * srate) # idx_qrs_end - idx_r #int(0.1*srate)
st_end = int(0.25 * srate)
avg_spect = np.mean(
spect[st_start:st_end, :],
axis=0) # between 100 (50) and 250 ms (125) from rpeak
avg_alt = avg_spect[int(0.5 * segbeats)]
# cum_spect_noise = cum_spect[int(0.4*segbeats):int(0.46*segbeats)] # noise level: 0.44-0.49 cycles / beat
avg_spect_noise = avg_spect[int(0.44 * segbeats):int(
0.49 * segbeats)] # noise level: 0.44-0.49 cycles / beat
# cum_spect_noise = cum_spect[int(0.33 * segbeats):int(0.48 * segbeats)] # noise level: 0.44-0.49 cycles / beat
avg_noise_avg = np.mean(avg_spect_noise)
avg_noise_std = np.std(avg_spect_noise)
# return avg beat
# avg_beat = np.mean(beats, axis=0)
for j in range(len(avg_beat)):
if len(ret_avg_beat['vals']) > r_list[seg_start + segbeats -
1] + j:
ret_avg_beat['vals'][r_list[seg_start + segbeats - 1] +
j] = avg_beat[j]
# print('avg alt {}, noise {}'.format(cum_alt, cum_noise_avg))
twar = 0
if avg_alt > avg_noise_avg:
twav = 1000 * (avg_alt - avg_noise_avg)**0.5
twar = (avg_alt - avg_noise_avg) / avg_noise_std
twavs.append(twav)
ret_twav.append({
'dt': r_list[seg_start + segbeats - 1] / srate,
'val': twav
})
twars.append(twar)
ret_twar.append({
'dt': r_list[seg_start + segbeats - 1] / srate,
'val': twar
})
# plt.figure(figsize=(30, 5))
# plt.plot(ecg_filt.tolist(), color='black', lw=1)
# plt.savefig('e:/{}_raw.pdf'.format(twar), bbox_inches="tight", pad_inches=0.5)
# plt.close()
#
# plt.figure(figsize=(10, 5))
# for i in range(len(beats)):
# c = 'red'
# if i % 2 == 0:
# c = 'blue'
# plt.plot(beats[i], color=c, lw=1)
# plt.savefig('e:/{}_ecg.pdf'.format(twar), bbox_inches="tight", pad_inches=0.5)
# plt.close()
#
# plt.figure(figsize=(10, 5))
# plt.plot(np.arange(1, 65) / 128, avg_spect[1:65], lw=1)
# plt.savefig('e:/{}_spect.pdf'.format(twar), bbox_inches="tight", pad_inches=0.5)
# plt.close()
dt_last = r_list[-1] / srate - cfg['overlap']
return [{
'srate': srate,
'vals': ecg_filt.tolist()
}, ret_avg_beat, ret_rpeak, ret_twav, ret_twar]
def run(inp, opt, cfg):
"""
calculate ppv from arterial waveform
:param art: arterial waveform
:return: max, min, upper envelope, lower envelope, respiratory rate, ppv
"""
global last_ppv
data = arr.interp_undefined(inp['pleth']['vals'])
srate = inp['pleth']['srate']
data = arr.resample_hz(data, srate, 100)
srate = 100
if len(data) < 30 * srate:
print('hr < 30')
return
# beat detection
minlist, maxlist = arr.detect_peaks(data, srate)
maxlist = maxlist[1:]
# beat lengths
beatlens = []
beats_128 = []
beats_128_valid = []
for i in range(0, len(minlist) - 1):
beatlen = minlist[i + 1] - minlist[i] # in samps
if not 30 < beatlen < 300:
beats_128.append(None)
continue
pp = data[maxlist[i]] - data[minlist[i]] # pulse pressure
if not 20 < pp < 100:
beats_128.append(None)
continue
beatlens.append(beatlen)
beat = data[minlist[i]:minlist[i + 1]]
resampled = arr.resample(beat, 128)
beats_128.append(resampled)
beats_128_valid.append(resampled)
if not beats_128_valid:
return
avgbeat = np.array(beats_128_valid).mean(axis=0)
meanlen = np.mean(beatlens)
stdlen = np.std(beatlens)
if stdlen > meanlen * 0.2: # irregular rhythm
return
# remove beats with correlation < 0.9
pulse_vals = []
for i in range(0, len(minlist) - 1):
if not beats_128[i]:
continue
if np.corrcoef(avgbeat, beats_128[i])[0, 1] < 0.9:
continue
pp = data[maxlist[i]] - data[minlist[i]] # pulse pressure
pulse_vals.append({'dt': minlist[i] / srate, 'val': pp})
# estimates the upper env(n) and lower env(n) envelopes
xa = np.array([data[idx] for idx in minlist])
lower_env = np.array([0.0] * len(data))
for i in range(len(data)):
be = np.array([b((i - idx) / (0.2 * srate)) for idx in minlist])
s = sum(be)
if s != 0:
lower_env[i] = np.dot(xa, be) / s
xb = np.array([data[idx] for idx in maxlist])
upper_env = np.array([0.0] * len(data))
for i in range(len(data)):
be = np.array([b((i - idx) / (0.2 * srate)) for idx in maxlist])
s = sum(be)
if s != 0:
upper_env[i] = np.dot(xb, be) / s
pulse_env = upper_env - lower_env
pulse_env[pulse_env < 0.0] = 0.0
# estimates resp rate
rr = arr.estimate_resp_rate(pulse_env, srate)
# split by respiration
nsamp_in_breath = int(srate * 60 / rr)
m = int(len(data) / nsamp_in_breath) # m segments exist
raw_pps = []
pps = []
for ibreath in np.arange(0, m - 1, 0.5):
pps_breath = []
for ppe in pulse_vals:
if ibreath * nsamp_in_breath < ppe['dt'] * srate < (
ibreath + 1) * nsamp_in_breath:
pps_breath.append(ppe['val'])
if len(pps_breath) < 4:
continue
pp_min = min(pps_breath)
pp_max = max(pps_breath)
ppv = 2 * (pp_max - pp_min) / (pp_max + pp_min) * 100 # estimate
if not 0 < ppv < 50:
continue
# raw_pps.append({'dt': (ibreath * nsamp_in_breath) / srate, 'val': pp})
#
# kalman filter
if last_ppv == 0: # first time
last_ppv = ppv
elif abs(last_ppv - ppv) <= 1.0:
ppv = last_ppv
elif abs(last_ppv - ppv) <= 25.0: # ppv cannot be changed abruptly
ppv = (ppv + last_ppv) * 0.5
last_ppv = ppv
else:
continue # no update
pps.append({
'dt': ((ibreath + 1) * nsamp_in_breath) / srate,
'val': int(ppv)
})
return [pps, pulse_vals, [{'dt': cfg['interval'], 'val': rr}]]
def run(inp, opt, cfg):
ecg_data = arr.interp_undefined(inp['ecg']['vals'])
ecg_srate = inp['ecg']['srate']
pleth_data = arr.interp_undefined(inp['pleth']['vals'])
pleth_srate = inp['pleth']['srate']
pleth_data = arr.band_pass(pleth_data, pleth_srate, 0.5, 15)
ecg_rlist = arr.detect_qrs(ecg_data, ecg_srate)
pleth_minlist, pleth_maxlist = arr.detect_peaks(pleth_data, pleth_srate)
dpleth = np.diff(pleth_data)
pleth_dmaxlist = [
] # index of the maximum slope between peak and nadir in pleth
for i in range(len(pleth_minlist)): # maxlist is one less than minlist
dmax_idx = arr.max_idx(dpleth, pleth_minlist[i], pleth_maxlist[i + 1])
pleth_dmaxlist.append(dmax_idx)
pttmax_list = []
pttmin_list = []
pttdmax_list = []
for i in range(len(ecg_rlist) - 1):
if len(pleth_minlist) == 0:
continue
if len(pleth_maxlist) == 0:
continue
rpeak_dt = ecg_rlist[i] / ecg_srate
rpeak_dt_next = ecg_rlist[i + 1] / ecg_srate
if rpeak_dt < cfg['overlap']:
continue
# find first min in pleth after rpeak_dt in ecg
found_minidx = 0
for minidx in pleth_minlist:
if minidx > rpeak_dt * pleth_srate:
found_minidx = minidx
break
elif minidx > rpeak_dt_next * pleth_srate:
break
if found_minidx == 0:
continue
# find first dmax in pleth after rpeak_dt in ecg
found_dmaxidx = 0
for dmaxidx in pleth_dmaxlist:
if dmaxidx > rpeak_dt * pleth_srate:
found_dmaxidx = dmaxidx
break
elif dmaxidx > rpeak_dt_next * pleth_srate:
break
if found_dmaxidx == 0:
continue
# find first dmax in pleth after rpeak_dt in ecg
found_maxidx = 0
for maxidx in pleth_maxlist:
if maxidx > rpeak_dt * pleth_srate:
found_maxidx = maxidx
break
elif maxidx > rpeak_dt_next * pleth_srate:
break
if found_maxidx == 0:
continue
max_dt = found_maxidx / pleth_srate
if max_dt > cfg['interval']:
continue
min_dt = found_minidx / pleth_srate
dmax_dt = found_dmaxidx / pleth_srate
pttmax_list.append({'dt': max_dt, 'val': (max_dt - rpeak_dt) * 1000})
pttdmax_list.append({
'dt': dmax_dt,
'val': (dmax_dt - rpeak_dt) * 1000
})
pttmin_list.append({'dt': min_dt, 'val': (min_dt - rpeak_dt) * 1000})
return [
pttmin_list, pttdmax_list,
arr.get_samples(ecg_data, ecg_srate, ecg_rlist), pttmax_list
]
def run(inp, opt, cfg):
"""
calculate ppv from arterial waveform
:param art: arterial waveform
:return: max, min, upper envelope, lower envelope, respiratory rate, ppv
"""
data = arr.interp_undefined(inp['pleth']['vals'])
srate = inp['pleth']['srate']
data = arr.resample_hz(data, srate, 100)
srate = 100
if len(data) < 30 * srate:
return [{}, {}, {}, {}, {}, [], []]
minlist, maxlist = arr.detect_peaks(data, srate)
maxlist = maxlist[1:]
# estimates the upper ue(n) and lower le(n) envelopes
xa = np.array([data[idx] for idx in minlist])
le = np.array([0] * len(data))
for i in range(len(data)):
be = np.array([b((i - idx) / (0.2 * srate)) for idx in minlist])
s = sum(be)
if s != 0:
le[i] = np.dot(xa, be) / s
xb = np.array([data[idx] for idx in maxlist])
ue = np.array([0] * len(data))
for i in range(len(data)):
be = np.array([b((i - idx) / (0.2 * srate)) for idx in maxlist])
s = sum(be)
if s != 0:
ue[i] = np.dot(xb, be) / s
re = ue - le
re[re < 0] = 0
# estimates resp rate
rr = arr.estimate_resp_rate(re, srate)
# split by respiration
nsamp_in_breath = int(srate * 60 / rr)
m = int(len(data) / nsamp_in_breath) # m segments exist
pps = []
for i in range(m - 1):
imax = arr.max_idx(re, i * nsamp_in_breath, (i+2) * nsamp_in_breath) # 50% overlapping
imin = arr.min_idx(re, i * nsamp_in_breath, (i+2) * nsamp_in_breath)
ppmax = re[imax]
ppmin = re[imin]
ppe = 2 * (ppmax - ppmin) / (ppmax + ppmin) * 100 # estimate
if ppe > 50 or ppe < 0:
continue
pp = cfg['pp']
if pp == 0:
pp = ppe
err = abs(ppe - pp)
if err < 1:
pp = ppe
elif err < 25:
pp = (pp + ppe) / 2
else:
pass # dont update
cfg['pp'] = pp
pps.append({'dt': (i * nsamp_in_breath) / srate, 'val': pp})
return [
[{'dt': cfg['interval'], 'val': rr}],
pps
]
def run(inp, opt, cfg):
"""
calculate ppv from arterial waveform
:param art: arterial waveform
:return: max, min, upper envelope, lower envelope, respiratory rate, ppv
"""
global last_ppv, last_spv
data = arr.interp_undefined(inp['ART']['vals'])
srate = inp['ART']['srate']
data = arr.resample_hz(data, srate, 100)
srate = 100
if len(data) < 30 * srate:
print('hr < 30')
return
# beat detection
minlist, maxlist = arr.detect_peaks(data, srate)
maxlist = maxlist[1:]
# beat lengths
beatlens = []
beats_128 = []
beats_128_valid = []
for i in range(0, len(minlist) - 1):
beatlen = minlist[i + 1] - minlist[i] # in samps
if not 30 < beatlen < 300:
beats_128.append(None)
continue
pp = data[maxlist[i]] - data[minlist[i]] # pulse pressure
if not 20 < pp < 100:
beats_128.append(None)
continue
beatlens.append(beatlen)
beat = data[minlist[i]:minlist[i + 1]]
resampled = arr.resample(beat, 128)
beats_128.append(resampled)
beats_128_valid.append(resampled)
if not beats_128_valid:
return
avgbeat = np.array(beats_128_valid).mean(axis=0)
meanlen = np.mean(beatlens)
stdlen = np.std(beatlens)
if stdlen > meanlen * 0.2: # irregular rhythm
return
# remove beats with correlation < 0.9
pp_vals = []
sp_vals = []
for i in range(0, len(minlist) - 1):
if beats_128[i] is None or not len(beats_128[i]):
continue
if np.corrcoef(avgbeat, beats_128[i])[0, 1] < 0.9:
continue
pp = data[maxlist[i]] - data[minlist[i]] # pulse pressure
sp = data[maxlist[i]]
pp_vals.append({'dt': minlist[i] / srate, 'val': pp})
sp_vals.append({'dt': minlist[i] / srate, 'val': sp})
dtstart = time.time()
# estimates resp rate
# upper env
idx_start = max(min(minlist), min(maxlist))
idx_end = min(max(minlist), max(maxlist))
xa = scipy.interpolate.CubicSpline(
maxlist, [data[idx] for idx in maxlist])(np.arange(idx_start, idx_end))
# lower env
xb = scipy.interpolate.CubicSpline(
minlist, [data[idx] for idx in minlist])(np.arange(idx_start, idx_end))
rr = arr.estimate_resp_rate(xa - xb, srate)
dtend = time.time()
#print('rr {}'.format(rr))
# split by respiration
nsamp_in_breath = int(srate * 60 / rr)
m = int(len(data) / nsamp_in_breath) # m segments exist
raw_pps = []
raw_sps = []
ppvs = []
spvs = []
for ibreath in np.arange(0, m - 1, 0.5):
pps_breath = []
sps_breath = []
for ppe in pp_vals:
if ibreath * nsamp_in_breath < ppe['dt'] * srate < (
ibreath + 1) * nsamp_in_breath:
pps_breath.append(ppe['val'])
for spe in sp_vals:
if ibreath * nsamp_in_breath < spe['dt'] * srate < (
ibreath + 1) * nsamp_in_breath:
sps_breath.append(spe['val'])
if len(pps_breath) < 4:
continue
if len(sps_breath) < 4:
continue
pp_min = min(pps_breath)
pp_max = max(pps_breath)
sp_min = min(sps_breath)
sp_max = max(sps_breath)
ppv = (pp_max - pp_min) / (pp_max + pp_min) * 200
if not 0 < ppv < 50:
continue
spv = (sp_max - sp_min) / (sp_max + sp_min) * 200
if not 0 < spv < 50:
continue
# kalman filter
if last_ppv == 0: # first time
last_ppv = ppv
elif abs(last_ppv - ppv) <= 1.0:
ppv = last_ppv
elif abs(last_ppv - ppv) <= 25.0: # ppv cannot be changed abruptly
ppv = (ppv + last_ppv) * 0.5
last_ppv = ppv
else:
continue
if last_spv == 0: # first time
last_spv = spv
elif abs(last_spv - spv) <= 1.0:
spv = last_spv
elif abs(last_spv - spv) <= 25.0: # ppv cannot be changed abruptly
spv = (spv + last_spv) * 0.5
last_spv = spv
else:
continue
ppvs.append(ppv)
spvs.append(spv)
median_ppv = np.median(ppvs)
median_spv = np.median(spvs)
return [[{
'dt': cfg['interval'],
'val': median_ppv
}], [{
'dt': cfg['interval'],
'val': median_spv
}], [{
'dt': cfg['interval'],
'val': rr
}]]
def run(inp, opt, cfg):
"""
calculate svv from arterial waveform
:param art: arterial waveform
:return: max, min, upper envelope, lower envelope, respiratory rate, ppv
"""
data = arr.interp_undefined(inp['art1']['vals'])
srate = inp['art1']['srate']
data = arr.resample_hz(data, srate, 100)
srate = 100
if len(data) < 30 * srate:
return [[], [], [], []]
minlist, maxlist = arr.detect_peaks(data, srate)
maxlist = maxlist[1:] # make the same length
# calculate each beat's std and put it at the peak time
stds = []
lzs = []
for i in range(len(minlist) - 1):
maxidx = maxlist[i]
beat = data[minlist[i]:minlist[i + 1]]
if max(beat) - min(beat) < 20:
continue
s = np.std(beat)
stds.append({'dt': maxidx / srate, 'val': s})
sbp = np.max(beat)
dbp = beat[0]
lz = (sbp - dbp) / (sbp + dbp) # 0.1~0.3
lzs.append({'dt': maxidx / srate, 'val': lz})
# estimates resp rate
rr = np.median([o['val'] for o in inp['vent_rr']])
if not rr > 1:
return [[], [], [], []]
# split by respiration
nsamp_in_breath = int(srate * 60 / rr)
m = int(len(data) / nsamp_in_breath) # m segments exist
# std
svv_stds = []
for i in range(m - 1): # 50% overlapping
this_breath_stds = []
for j in range(len(stds)):
if i * nsamp_in_breath <= stds[j]['dt'] * srate < (
i + 2) * nsamp_in_breath:
this_breath_stds.append(stds[j]['val'])
svmax = np.max(this_breath_stds)
svmin = np.min(this_breath_stds)
svv_stde = 2 * (svmax - svmin) * 100 / (svmax + svmin) # estimate
if svv_stde > 40 or svv_stde < 0:
continue
svv_stds.append(svv_stde)
svv_stde = np.median(svv_stds)
if svv_stde < 0:
svv_stde = 0
svv_std = cfg['svv_std']
if svv_std == 0 or svv_std is None:
svv_std = svv_stde
err = abs(svv_stde - svv_std)
if err < 5:
svv_std = svv_stde
elif err < 25:
svv_std = (svv_std + svv_stde) / 2
else:
pass # dont update
cfg['svv_std'] = svv_std
# lz
svv_lzs = []
for i in range(m - 1): # 50% overlapping
this_breath_lzs = []
for j in range(len(lzs)):
if i * nsamp_in_breath <= lzs[j]['dt'] * srate < (
i + 2) * nsamp_in_breath:
this_breath_lzs.append(lzs[j]['val'])
svmax = np.max(this_breath_lzs)
svmin = np.min(this_breath_lzs)
svv_lze = 2 * (svmax - svmin) * 100 / (svmax + svmin) # estimate
if svv_lze > 40 or svv_lze < 0:
continue
svv_lzs.append(svv_lze)
svv_lze = np.median(svv_lzs)
if svv_lze < 0:
svv_lze = 0
svv_lz = cfg['svv_lz']
if svv_lz == 0 or svv_lz is None:
svv_lz = svv_lze
err = abs(svv_lze - svv_lz)
if err < 5:
svv_lz = svv_lze
elif err < 25:
svv_lz = (svv_lz + svv_lze) / 2
else:
pass # dont update
cfg['svv_lz'] = svv_lz
return [
stds, [{
'dt': cfg['interval'],
'val': svv_std
}], lzs, [{
'dt': cfg['interval'],
'val': svv_lz
}]
]
def run(inp, opt, cfg):
data = arr.interp_undefined(inp['ecg']['vals'])
srate = inp['ecg']['srate']
min_hr = 40 # min bpm
max_hr = 200 # max bpm
min_qrs = 0.04 # min qslist duration
max_qrs = 0.2 # max qslist duration
min_umv = 0.2 # min UmV of R,S peaks
min_pq = 0.07 # min PQ duration
max_pq = 0.20 # max PQ duration
min_qt = 0.21 # min QT duration
max_qt = 0.48 # max QT duration
pfreq = 9.0 # cwt Hz for pidx wave
tfreq = 2.5 # cwt Hz for tidx wave
min_sq = (60.0 / max_hr) - max_qrs # from s to next q
if min_sq * srate <= 0:
min_sq = 0.1
max_hr = int(60.0 / (max_qrs + min_sq))
# denoised ecg
depth = int(math.ceil(np.log2(srate / 0.8))) - 1
ad = pywt.wavedec(data, 'db2', level=depth)
ad[0].fill(0) # low frequency approx -> 0
ecg_denoised = pywt.waverec(ad, 'db2')
# interpolation filter
inter1 = pywt.Wavelet('inter1', filter_bank=orthfilt([0.25, 0.5, 0.25]))
# qrs augmented ecg
sig = cwt(data, srate, 'gaus1', 13) # 13 Hz gaus convolution
depth = int(math.ceil(np.log2(srate / 23))) - 2
ad = pywt.wavedec(sig, inter1, level=depth)
for level in range(depth): # remove [0-30Hz]
wsize = int(2 * srate / (2**(level + 1))) # 2 sec window
denoise(ad[depth - level],
wsize) # Remove less than 30 hz from all detail
ad[0].fill(0) # most lowest frequency approx -> 0
ecg_qrs = pywt.waverec(ad, inter1)
# start parsing
qslist = [] # qrs list [startqrs, endqrs, startqrs, endqrs, ...]
vpclist = [] # abnormal beat
# save greater than 0 after min_sq
prev_zero = 0
ipos = 0
while ipos < len(ecg_qrs) - int(max_qrs * srate):
if ecg_qrs[ipos] == 0:
prev_zero += 1
else:
if prev_zero > min_sq * srate:
iend = ipos + int(
max_qrs *
srate) # find the position of the end of the current qrs
while iend > ipos:
if ecg_qrs[iend] != 0:
break
iend -= 1
# Check if it is the minimum length or if there is a pause
if ipos + min_qrs * srate > iend or np.any(
ecg_qrs[iend + 1:iend + 1 + int(min_sq * srate)]):
vpclist.append(ipos) # push vpc
else:
qslist.append(ipos)
qslist.append(iend)
ipos = iend
prev_zero = 0
ipos += 1
# qlist = [qslist[i] for i in range(0, len(qslist), 2)]
complist = []
for n in range(int(len(qslist) / 2)):
start_qrs = qslist[n * 2]
end_qrs = qslist[n * 2 + 1]
qidx = -1
ridx = arr.max_idx(ecg_denoised, start_qrs, end_qrs)
if ecg_denoised[ridx] < min_umv:
ridx = -1
sidx = arr.min_idx(ecg_denoised, start_qrs, end_qrs)
if -ecg_denoised[sidx] < min_umv:
sidx = -1
# ridxpeak > 0mV sidxpeak < 0mV
if ridx != -1 and sidx != -1:
if sidx < ridx: # check for sidx
if ecg_denoised[ridx] > -ecg_denoised[sidx]:
qidx = sidx
sidx = arr.min_idx(ecg_denoised, ridx, end_qrs + 1)
if sidx == ridx or sidx == end_qrs or abs(
ecg_denoised[end_qrs] - ecg_denoised[sidx]) < 0.05:
sidx = -1
else: # check for qidx
qidx = arr.min_idx(ecg_denoised, start_qrs, ridx + 1)
if qidx == ridx or qidx == start_qrs or abs(
ecg_denoised[start_qrs] - ecg_denoised[qidx]) < 0.05:
qidx = -1
elif sidx != -1: # only sidx --> Find small r if only sidx detected in rsidx large tidx lead
ridx = arr.max_idx(ecg_denoised, start_qrs, sidx + 1)
if ridx == sidx or ridx == start_qrs or abs(
ecg_denoised[start_qrs] - ecg_denoised[ridx]) < 0.05:
ridx = -1
elif ridx != -1: # only ridx --> Find small q,s
qidx = arr.min_idx(ecg_denoised, start_qrs, ridx + 1)
if qidx == ridx or qidx == start_qrs or abs(
ecg_denoised[start_qrs] - ecg_denoised[qidx]) < 0.05:
qidx = -1
sidx = arr.min_idx(ecg_denoised, ridx, end_qrs + 1)
if sidx == ridx or sidx == end_qrs or abs(
ecg_denoised[end_qrs] - ecg_denoised[sidx]) < 0.05:
sidx = -1
else:
vpclist.append(start_qrs)
continue
o = {'q': qslist[n * 2], 's': qslist[n * 2 + 1]} # always exists
if qidx != -1:
o['q'] = qidx
if ridx != -1:
o['r'] = ridx
if sidx != -1:
o['s'] = sidx
complist.append(o)
# for each QRS --> find tidx and pidx wave
for n in range(len(complist) - 1):
pree = complist[n]['q']
nows = complist[n]['s']
nowe = complist[n + 1]['q']
size = nowe - nows # s-q interval
size = int(min(size, srate * max_qt - (nows - pree)))
rr = (nowe - pree) / srate
if (60.0 / rr < min_hr) or (60.0 / rr > max_hr - 20):
continue
# all are in this
block = [data[nows + i] for i in range(size)]
ecg_qrs = cwt(block, srate, 'gaus1', tfreq)
tidx1 = arr.min_idx(ecg_qrs) + nows
tidx2 = arr.max_idx(ecg_qrs) + nows
if tidx1 > tidx2:
tidx1, tidx2 = tidx2, tidx1
# additional constraints on [tidx1 tidx tidx2] duration, symmetry, QT interval
ist = False
if ecg_qrs[tidx1 - nows] < 0 < ecg_qrs[tidx2 - nows]:
ist = True
elif ecg_qrs[tidx1 - nows] > 0 > ecg_qrs[tidx2 - nows]:
ist = True
if ist:
if (
tidx2 - tidx1
) >= 0.09 * srate: # and (tidx2-tidx1)<=0.24 * srate) #check for tidx wave duration
ist = True # QT interval = .4 * sqrt(RR)
if min_qt * srate <= (tidx2 - pree) <= max_qt * srate:
ist = True
else:
ist = False
else:
ist = False
if ist:
tidx = 0 # zero crossing
sign = (ecg_qrs[tidx1 - nows] >= 0)
for i in range(tidx1 - nows, tidx2 - nows):
if sign == (ecg_qrs[i] >= 0):
continue
tidx = i + nows
break
# check for tidx wave symetry
if tidx2 - tidx < tidx - tidx1:
ratio = (tidx2 - tidx) / (tidx - tidx1)
else:
ratio = (tidx - tidx1) / (tidx2 - tidx)
if ratio < 0.4:
ist = False
if ist:
tmin = arr.min_idx(data, tidx1, tidx2)
tmax = arr.max_idx(data, tidx1, tidx2)
# find the most nearest values from 0-cross, tmin, tmax
tidx = arr.find_nearest((tidx, tmin, tmax), (tidx2 + tidx1) / 2)
complist[n]['(t'] = tidx1
complist[n]['t'] = tidx
complist[n]['t)'] = tidx2
# search for P-WAVE
size = nowe - nows # s-q interval
size = int(min(size, srate * max_pq))
if ist:
if tidx2 > nowe - size - int(
0.04 *
srate): # isp wnd far from Twave at least on 0.04 sec
size -= tidx2 - (nowe - size - int(0.04 * srate))
nskip = (nowe - nows) - size
if size <= 0.03 * srate:
continue # impresize QRS begin detection
block = [data[nows + nskip + i] for i in range(size)]
ecg_qrs = cwt(block, srate, 'gaus1', pfreq)
p1 = arr.min_idx(ecg_qrs) + nows + nskip
p2 = arr.max_idx(ecg_qrs) + nows + nskip
if p1 > p2:
p1, p2 = p2, p1
# additional constraints on [p1 pidx p2] duration, symmetry, PQ interval
isp = False
if ecg_qrs[p1 - nows - nskip] < 0 < ecg_qrs[p2 - nows - nskip]:
isp = True
elif ecg_qrs[p1 - nows - nskip] > 0 > ecg_qrs[p2 - nows - nskip]:
isp = True
if isp:
if 0.03 * srate <= (
p2 - p1
) <= 0.15 * srate: # check for pidx wave duration 9Hz0.03 5Hz0.05
isp = (min_pq * srate <= (nowe - p1) <= max_pq * srate
) # PQ interval = [0.07 - 0.12,0.20]
else:
isp = False
if not isp:
continue
pidx = 0 # zero crossing
sign = (ecg_qrs[p1 - nows - nskip] >= 0)
for i in range(p1 - nows - nskip, p2 - nows - nskip):
if sign == (ecg_qrs[i] >= 0):
continue
pidx = i + nows + nskip
break
# check for pidx wave symetry
if p2 - pidx < pidx - p1:
ratio = (p2 - pidx) / (pidx - p1)
else:
ratio = (pidx - p1) / (p2 - pidx)
if ratio < 0.4:
isp = False # not a p wave
if isp:
complist[n]['(p'] = p1
complist[n]['p'] = pidx
complist[n]['p)'] = p2
# add annotation
ret_ann = []
for n in range(len(complist)):
for k, v in complist[n].items():
if k == 'q' and abs(ecg_denoised[v]) > 0.5:
k = 'Q'
elif k == 'r' and abs(ecg_denoised[v]) > 0.5:
k = 'R'
elif k == 's' and abs(ecg_denoised[v]) > 0.5:
k = 'S'
elif k == '(t':
k = '(T'
elif k == 't':
k = 'T'
elif k == 't)':
k = 'T)'
elif k == '(p':
k = '(P'
elif k == 'p':
k = 'P'
elif k == 'p)':
k = 'P)'
ret_ann.append({"dt": v / srate, "val": k})
for n in range(len(vpclist)):
ret_ann.append({"dt": vpclist[n] / srate, "val": 'A'})
return [ret_ann]
