def sqi2(self):
"""linear resampling (between two fiducials up to length L, correlation)"""
slicez, islicez = self.slices(method='variable')
L = len(self.template)
corrs = np.array([(cross_corr(sig_resample(sl, L), self.template)
if len(sl) else 0.0) for sl in slicez])
corrs = np.clip(corrs, a_min=0.0, a_max=1.0)
return corrs
def sqi3(self):
"""DTW resampling (resampling to length L and correlation)"""
slicez, islicez = self.slices(method='variable')
corrs = np.array([
cross_corr(*self.dtw_resample(sl)) if len(sl) else 0.0
for sl in slicez
])
corrs = np.clip(corrs, a_min=0.0, a_max=1.0)
return corrs
def sqi1(self):
"""direct matching (fiducial + length L template correlation)"""
# nb. slight difference: we are centering the window on the beat, while Li et al
slicez, islicez = self.slices(method='fixed')
corrs = np.array([
cross_corr(sl, self.template) if len(sl) else 0.0 for sl in slicez
])
corrs = np.clip(corrs, a_min=0.0, a_max=1.0)
return corrs
def ecg_kept(raw):
sig0 = raw
fps = sig0.fps
# print 'slice.x shape=', sig0.slice(slice(int(0*fps), int(ECG_DURATION*fps))).x.shape
###
ecg = beatdet_ecg(sig0.slice(slice(int(0 * fps), int(ECG_DURATION * fps))))
ecg.lead = LEADS[ECG_LEAD]
# scaling for plot
pcs = np.array([np.percentile(ecg.x, 10), np.percentile(ecg.x, 90)])
pcs = (pcs - np.mean(pcs)) * 5.0 + np.mean(pcs)
###
slicez = sqi_slices(ecg, method='fixed', slice_front=0.5, slice_back=-0.5)
L = max([len(sl) for sl in slicez])
padded_slicez = np.array(
[sig_pad(sl, L, side='center', mode='constant') for sl in slicez])
###
perc = envelopes_perc_threshold(padded_slicez)
le, ue = envelopes_at_perc(padded_slicez, perc)
mb = np.median(padded_slicez, axis=0)
bpt = beat_penalty_threshold(le, ue, mb)
bps = np.array([beat_penalty(sl, le, ue, mb) for sl in padded_slicez])
bp_ok = bps < bpt
###
template = np.median(padded_slicez, axis=0)
corrs = np.array([cross_corr(sl, template) for sl in padded_slicez])
CORR_THRESHOLD = 0.8
corr_ok = corrs > CORR_THRESHOLD
###
# next beat is OK as well?
bp_ibi_ok = bp_ok & np.roll(bp_ok, -1)
corr_ibi_ok = corr_ok & np.roll(corr_ok, -1)
igood = np.where(bp_ibi_ok & corr_ibi_ok)[0]
def beat_template_1(self):
self.L = np.median(np.diff(self.tbeats))
slicez_1, islicez_1 = self.slices(
method="fixed") #, hwin=int(self.L*self.fps/2.)))
self.ibis_good = islicez_1
template_1 = np.mean(slicez_1, axis=0)
#print 'template_1', template_1
slicez, islicez = self.slices(method="fixed", scrub=False)
corrs = np.array([cross_corr(sl, template_1) for sl in slicez])
self.slicez_1, self.islicez_1, self.template_1 = slicez_1, islicez_1, template_1
self.slicez, self.islicez, self.corrs = slicez, islicez, corrs
# TODO: penalize the correlation of overlong IBIs (since correlation only goes until the beat template ends)
# debug: put shape envelopes
self.env_min = sqi_copy_to_idxs(self.s_min, len(self.x),
self.ibeats.astype(int),
[sqi_slice_norm(sl) for sl in slicez])
self.env_max = sqi_copy_to_idxs(self.s_max, len(self.x),
self.ibeats.astype(int),
[sqi_slice_norm(sl) for sl in slicez])
def beat_detect(self, debug=False, outlierthreshold=0.001):
# Heuristics:
# * found enough beats
# * median ibi is in plausible range (or even: some percentile of ibis is in plausible range)
# * histogram of beat correlations is plausible (lots of good correlation)
#
# Good beats have positive correlation, e.g. rho > 0.5 with the median beat.
ecg = self.ecg
#ecg.x[:int(ecg.fps*5)] *= 0.0 # avoid the terrible swing
#
# Kim ECG beat detection
#
smoothsignal = np.array(ecg.x)
# kill outliers
mn, mx = np.min(smoothsignal), np.max(smoothsignal)
m = min(abs(mn), abs(mx))
N = 100
step = m / float(N)
for i in range(N):
n = len(np.where(smoothsignal < -m)[0]) + len(
np.where(smoothsignal > m)[0])
if n > outlierthreshold * len(smoothsignal):
break
m -= step
mn, mx = -m, m
smoothsignal[smoothsignal < mn] = mn
smoothsignal[smoothsignal > mx] = mx
smoothsignal[-10:] = 0 # extreme outlier in last few frames
# adjust distribution to the one Kim has optimized for
smoothsignal = (smoothsignal - np.mean(smoothsignal)
) / np.std(smoothsignal) * 0.148213 - 0.191034
loc, beattime = self.QRSdetection(smoothsignal,
ecg.fps,
ecg.t,
ftype=0)
loc = loc.flatten()
#
# check error Kim vs. localmax of R peaks
#
new_loc = localmax_climb(ecg.x, loc,
hwin=int(0.02 * ecg.fps)) # hwin = 20 ms
peak_errs = (new_loc - loc) / float(ecg.fps)
#print 'np.mean(peak_errs), np.std(peak_errs)', np.mean(peak_errs), np.std(peak_errs)
ibis = np.diff(loc / float(ecg.fps))
median_ibi = np.median(ibis)
#
# filter beats by cross-correlation with median beat
#
ecg_slices = np.array(
slices(ecg.x, loc, hwin=int(np.ceil(median_ibi * ecg.fps)) // 2))
# median value from each timepoint (not a single one of any of the beats)
median_beat = np.median(ecg_slices, axis=0)
if debug:
plt.plot(np.arange(len(median_beat)) / float(ecg.fps), median_beat)
plt.title('median ECG beat')
cross_corrs = [cross_corr(sl, median_beat) for sl in ecg_slices]
spectrum_ok = np.array(
[self.slice_good(sl, median_beat) for sl in ecg_slices])
ccs_ok = np.array(cross_corrs) > NoisyECG.GOOD_BEAT_THRESHOLD
good_loc_idxs = np.where(ccs_ok & spectrum_ok)[0]
if debug:
[
plt.plot(
np.arange(len(ecg_slices[i])) / float(ecg.fps),
ecg_slices[i]) for i in range(1, len(ecg_slices))
if i in good_loc_idxs
]
plt.title('all good ECG beats with rho > {:.2f}'.format(
NoisyECG.GOOD_BEAT_THRESHOLD))
plt.show()
beat_idxs = loc[good_loc_idxs]
beat_times = beattime[good_loc_idxs]
self._beattime, self._cross_corrs = beattime, cross_corrs
if debug:
self.debug_plot()
#self.cross_corrs = np.array(cross_corrs)[good_loc_idxs]
self.median_ibi = median_ibi
self.good_beat_fraction = float(len(good_loc_idxs)) / len(cross_corrs)
return beat_idxs, beat_times