def engzee_segmenter(signal=None, sampling_rate=1000., threshold=0.48):
# check inputs
if signal is None:
raise TypeError("Please specify an input signal.")
# algorithm parameters
changeM = int(0.75 * sampling_rate)
Miterate = int(1.75 * sampling_rate)
v250ms = int(0.25 * sampling_rate)
v1200ms = int(1.2 * sampling_rate)
v1500ms = int(1.5 * sampling_rate)
v180ms = int(0.18 * sampling_rate)
p10ms = int(np.ceil(0.01 * sampling_rate))
p20ms = int(np.ceil(0.02 * sampling_rate))
err_kill = int(0.01 * sampling_rate)
inc = 1
mmth = threshold
mmp = 0.2
# Differentiator (1)
y1 = [signal[i] - signal[i - 4] for i in range(4, len(signal))]
# Low pass filter (2)
c = [1, 4, 6, 4, 1, -1, -4, -6, -4, -1]
y2 = np.array([np.dot(c, y1[n - 9:n + 1]) for n in range(9, len(y1))])
y2_len = len(y2)
# vars
MM = mmth * max(y2[:Miterate]) * np.ones(3)
MMidx = 0
Th = np.mean(MM)
NN = mmp * min(y2[:Miterate]) * np.ones(2)
NNidx = 0
ThNew = np.mean(NN)
update = False
nthfpluss = []
rpeaks = []
# Find nthf+ point
while True:
# If a previous intersection was found, continue the analysis from there
if update:
if inc * changeM + Miterate < y2_len:
a = (inc - 1) * changeM
b = inc * changeM + Miterate
Mnew = mmth * max(y2[a:b])
Nnew = mmp * min(y2[a:b])
elif y2_len - (inc - 1) * changeM > v1500ms:
a = (inc - 1) * changeM
Mnew = mmth * max(y2[a:])
Nnew = mmp * min(y2[a:])
if len(y2) - inc * changeM > Miterate:
MM[MMidx] = Mnew if Mnew <= 1.5 * MM[MMidx - 1] else 1.1 * MM[
MMidx - 1]
NN[NNidx] = Nnew if abs(Nnew) <= 1.5 * abs(
NN[NNidx - 1]) else 1.1 * NN[NNidx - 1]
MMidx = np.mod(MMidx + 1, len(MM))
NNidx = np.mod(NNidx + 1, len(NN))
Th = np.mean(MM)
ThNew = np.mean(NN)
inc += 1
update = False
if nthfpluss:
lastp = nthfpluss[-1] + 1
if lastp < (inc - 1) * changeM:
lastp = (inc - 1) * changeM
y22 = y2[lastp:inc * changeM + err_kill]
# find intersection with Th
try:
nthfplus = np.intersect1d(
np.nonzero(y22 > Th)[0],
np.nonzero(y22 < Th)[0] - 1)[0]
except IndexError:
if inc * changeM > len(y2):
break
else:
update = True
continue
# adjust index
nthfplus += int(lastp)
# if a previous R peak was found:
if rpeaks:
# check if intersection is within the 200-1200 ms interval. Modification: 300 ms -> 200 bpm
if nthfplus - rpeaks[-1] > v250ms and nthfplus - rpeaks[
-1] < v1200ms:
pass
# if new intersection is within the <200ms interval, skip it. Modification: 300 ms -> 200 bpm
elif nthfplus - rpeaks[-1] < v250ms:
nthfpluss += [nthfplus]
continue
# no previous intersection, find the first one
else:
try:
aux = np.nonzero(
y2[(inc - 1) * changeM:inc * changeM + err_kill] > Th)[0]
bux = np.nonzero(y2[
(inc - 1) * changeM:inc * changeM + err_kill] < Th)[0] - 1
nthfplus = int(
(inc - 1) * changeM) + np.intersect1d(aux, bux)[0]
except IndexError:
if inc * changeM > len(y2):
break
else:
update = True
continue
nthfpluss += [nthfplus]
# Define 160ms search region
windowW = np.arange(nthfplus, nthfplus + v180ms)
# Check if the condition y2[n] < Th holds for a specified
# number of consecutive points (experimentally we found this number to be at least 10 points)"
i, f = windowW[0], windowW[-1] if windowW[-1] < len(y2) else -1
hold_points = np.diff(np.nonzero(y2[i:f] < ThNew)[0])
cont = 0
for hp in hold_points:
if hp == 1:
cont += 1
if cont == p10ms - 1: # -1 is because diff eats a sample
max_shift = p20ms # looks for X's max a bit to the right
if nthfpluss[-1] > max_shift:
rpeaks += [
np.argmax(signal[i - max_shift:f]) + i - max_shift
]
else:
rpeaks += [np.argmax(signal[i:f]) + i]
break
else:
cont = 0
rpeaks = sorted(list(set(rpeaks)))
rpeaks = np.array(rpeaks, dtype='int')
return utils.ReturnTuple((rpeaks, ), ('rpeaks', ))
def hamilton_segmenter(signal=None, sampling_rate=1000.):
if signal is None:
raise TypeError("Please specify an input signal.")
sampling_rate = float(sampling_rate)
length = len(signal)
dur = length / sampling_rate
# algorithm parameters
v1s = int(1. * sampling_rate)
v100ms = int(0.1 * sampling_rate)
TH_elapsed = np.ceil(0.36 * sampling_rate)
sm_size = int(0.08 * sampling_rate)
init_ecg = 8 # seconds for initialization
if dur < init_ecg:
init_ecg = int(dur)
# filtering
filtered, _, _ = st.filter_signal(signal=signal,
ftype='butter',
band='lowpass',
order=4,
frequency=25.,
sampling_rate=sampling_rate)
filtered, _, _ = st.filter_signal(signal=filtered,
ftype='butter',
band='highpass',
order=4,
frequency=3.,
sampling_rate=sampling_rate)
# diff
dx = np.abs(np.diff(filtered, 1) * sampling_rate)
# smoothing
dx, _ = st.smoother(signal=dx, kernel='hamming', size=sm_size, mirror=True)
# buffers
qrspeakbuffer = np.zeros(init_ecg)
noisepeakbuffer = np.zeros(init_ecg)
peak_idx_test = np.zeros(init_ecg)
noise_idx = np.zeros(init_ecg)
rrinterval = sampling_rate * np.ones(init_ecg)
a, b = 0, v1s
all_peaks, _ = st.find_extrema(signal=dx, mode='max')
for i in range(init_ecg):
peaks, values = st.find_extrema(signal=dx[a:b], mode='max')
try:
ind = np.argmax(values)
except ValueError:
pass
else:
# peak amplitude
qrspeakbuffer[i] = values[ind]
# peak location
peak_idx_test[i] = peaks[ind] + a
a += v1s
b += v1s
# thresholds
ANP = np.median(noisepeakbuffer)
AQRSP = np.median(qrspeakbuffer)
TH = 0.475
DT = ANP + TH * (AQRSP - ANP)
DT_vec = []
indexqrs = 0
indexnoise = 0
indexrr = 0
npeaks = 0
offset = 0
beats = []
# detection rules
# 1 - ignore all peaks that precede or follow larger peaks by less than 200ms
lim = int(np.ceil(0.2 * sampling_rate))
diff_nr = int(np.ceil(0.045 * sampling_rate))
bpsi, bpe = offset, 0
for f in all_peaks:
DT_vec += [DT]
# 1 - Checking if f-peak is larger than any peak following or preceding it by less than 200 ms
peak_cond = np.array(
(all_peaks > f - lim) * (all_peaks < f + lim) * (all_peaks != f))
peaks_within = all_peaks[peak_cond]
if peaks_within.any() and (max(dx[peaks_within]) > dx[f]):
continue
# 4 - If the peak is larger than the detection threshold call it a QRS complex, otherwise call it noise
if dx[f] > DT:
# 2 - look for both positive and negative slopes in raw signal
if f < diff_nr:
diff_now = np.diff(signal[0:f + diff_nr])
elif f + diff_nr >= len(signal):
diff_now = np.diff(signal[f - diff_nr:len(dx)])
else:
diff_now = np.diff(signal[f - diff_nr:f + diff_nr])
diff_signer = diff_now[diff_now > 0]
if len(diff_signer) == 0 or len(diff_signer) == len(diff_now):
continue
# RR INTERVALS
if npeaks > 0:
# 3 - in here we check point 3 of the Hamilton paper
# that is, we check whether our current peak is a valid R-peak.
prev_rpeak = beats[npeaks - 1]
elapsed = f - prev_rpeak
# if the previous peak was within 360 ms interval
if elapsed < TH_elapsed:
# check current and previous slopes
if prev_rpeak < diff_nr:
diff_prev = np.diff(signal[0:prev_rpeak + diff_nr])
elif prev_rpeak + diff_nr >= len(signal):
diff_prev = np.diff(signal[prev_rpeak -
diff_nr:len(dx)])
else:
diff_prev = np.diff(
signal[prev_rpeak - diff_nr:prev_rpeak + diff_nr])
slope_now = max(diff_now)
slope_prev = max(diff_prev)
if (slope_now < 0.5 * slope_prev):
# if current slope is smaller than half the previous one, then it is a T-wave
continue
if dx[f] < 3. * np.median(
qrspeakbuffer): # avoid retarded noise peaks
beats += [int(f) + bpsi]
else:
continue
if bpe == 0:
rrinterval[indexrr] = beats[npeaks] - beats[npeaks - 1]
indexrr += 1
if indexrr == init_ecg:
indexrr = 0
else:
if beats[npeaks] > beats[bpe - 1] + v100ms:
rrinterval[indexrr] = beats[npeaks] - beats[npeaks - 1]
indexrr += 1
if indexrr == init_ecg:
indexrr = 0
elif dx[f] < 3. * np.median(qrspeakbuffer):
beats += [int(f) + bpsi]
else:
continue
npeaks += 1
qrspeakbuffer[indexqrs] = dx[f]
peak_idx_test[indexqrs] = f
indexqrs += 1
if indexqrs == init_ecg:
indexqrs = 0
if dx[f] <= DT:
tf = f + bpsi
# RR interval median
RRM = np.median(rrinterval) # initial values are good?
if len(beats) >= 2:
elapsed = tf - beats[npeaks - 1]
if elapsed >= 1.5 * RRM and elapsed > TH_elapsed:
if dx[f] > 0.5 * DT:
beats += [int(f) + offset]
# RR INTERVALS
if npeaks > 0:
rrinterval[indexrr] = beats[npeaks] - beats[npeaks
- 1]
indexrr += 1
if indexrr == init_ecg:
indexrr = 0
npeaks += 1
qrspeakbuffer[indexqrs] = dx[f]
peak_idx_test[indexqrs] = f
indexqrs += 1
if indexqrs == init_ecg:
indexqrs = 0
else:
noisepeakbuffer[indexnoise] = dx[f]
noise_idx[indexnoise] = f
indexnoise += 1
if indexnoise == init_ecg:
indexnoise = 0
else:
noisepeakbuffer[indexnoise] = dx[f]
noise_idx[indexnoise] = f
indexnoise += 1
if indexnoise == init_ecg:
indexnoise = 0
# Update Detection Threshold
ANP = np.median(noisepeakbuffer)
AQRSP = np.median(qrspeakbuffer)
DT = ANP + 0.475 * (AQRSP - ANP)
beats = np.array(beats)
r_beats = []
thres_ch = 0.85
adjacency = 0.05 * sampling_rate
for i in beats:
error = [False, False]
if i - lim < 0:
window = signal[0:i + lim]
add = 0
elif i + lim >= length:
window = signal[i - lim:length]
add = i - lim
else:
window = signal[i - lim:i + lim]
add = i - lim
# meanval = np.mean(window)
w_peaks, _ = st.find_extrema(signal=window, mode='max')
w_negpeaks, _ = st.find_extrema(signal=window, mode='min')
zerdiffs = np.where(np.diff(window) == 0)[0]
w_peaks = np.concatenate((w_peaks, zerdiffs))
w_negpeaks = np.concatenate((w_negpeaks, zerdiffs))
pospeaks = sorted(zip(window[w_peaks], w_peaks), reverse=True)
negpeaks = sorted(zip(window[w_negpeaks], w_negpeaks))
try:
twopeaks = [pospeaks[0]]
except IndexError:
twopeaks = []
try:
twonegpeaks = [negpeaks[0]]
except IndexError:
twonegpeaks = []
# getting positive peaks
for i in range(len(pospeaks) - 1):
if abs(pospeaks[0][1] - pospeaks[i + 1][1]) > adjacency:
twopeaks.append(pospeaks[i + 1])
break
try:
posdiv = abs(twopeaks[0][0] - twopeaks[1][0])
except IndexError:
error[0] = True
# getting negative peaks
for i in range(len(negpeaks) - 1):
if abs(negpeaks[0][1] - negpeaks[i + 1][1]) > adjacency:
twonegpeaks.append(negpeaks[i + 1])
break
try:
negdiv = abs(twonegpeaks[0][0] - twonegpeaks[1][0])
except IndexError:
error[1] = True
# choosing type of R-peak
n_errors = sum(error)
try:
if not n_errors:
if posdiv > thres_ch * negdiv:
# pos noerr
r_beats.append(twopeaks[0][1] + add)
else:
# neg noerr
r_beats.append(twonegpeaks[0][1] + add)
elif n_errors == 2:
if abs(twopeaks[0][1]) > abs(twonegpeaks[0][1]):
# pos allerr
r_beats.append(twopeaks[0][1] + add)
else:
# neg allerr
r_beats.append(twonegpeaks[0][1] + add)
elif error[0]:
# pos poserr
r_beats.append(twopeaks[0][1] + add)
else:
# neg negerr
r_beats.append(twonegpeaks[0][1] + add)
except IndexError:
continue
rpeaks = sorted(list(set(r_beats)))
rpeaks = np.array(rpeaks, dtype='int')
return utils.ReturnTuple((rpeaks, ), ('rpeaks', ))
def ecg(signal=None, sampling_rate=300., show=True):
if signal is None:
raise TypeError("Please specify an input signal.")
# ensure numpy
signal = np.array(signal)
sampling_rate = float(sampling_rate)
# filter signal
order = int(0.3 * sampling_rate)
filtered, _, _ = st.filter_signal(signal=signal,
ftype='FIR',
band='bandpass',
order=order,
frequency=[3, 45],
sampling_rate=sampling_rate)
# segment
rpeaks, = hamilton_segmenter(signal=filtered, sampling_rate=sampling_rate)
# correct R-peak locations
rpeaks, = correct_rpeaks(signal=filtered,
rpeaks=rpeaks,
sampling_rate=sampling_rate,
tol=0.05)
# extract templates
templates, rpeaks = extract_heartbeats(signal=filtered,
rpeaks=rpeaks,
sampling_rate=sampling_rate,
before=0.2,
after=0.4)
# compute heart rate
hr_idx, hr = st.get_heart_rate(beats=rpeaks,
sampling_rate=sampling_rate,
smooth=True,
size=3)
# get time vectors
length = len(signal)
T = (length - 1) / sampling_rate
ts = np.linspace(0, T, length, endpoint=False)
ts_hr = ts[hr_idx]
ts_tmpl = np.linspace(-0.2, 0.4, templates.shape[1], endpoint=False)
# plot
if show:
plotting.plot_ecg(ts=ts,
raw=signal,
filtered=filtered,
rpeaks=rpeaks,
templates_ts=ts_tmpl,
templates=templates,
heart_rate_ts=ts_hr,
heart_rate=hr,
path=None,
show=True)
# output
args = (ts, filtered, rpeaks, ts_tmpl, templates, ts_hr, hr)
names = ('ts', 'filtered', 'rpeaks', 'templates_ts', 'templates',
'heart_rate_ts', 'heart_rate')
return utils.ReturnTuple(args, names)
def christov_segmenter(signal=None, sampling_rate=1000.):
# check inputs
if signal is None:
raise TypeError("Please specify an input signal.")
length = len(signal)
# algorithm parameters
v100ms = int(0.1 * sampling_rate)
v50ms = int(0.050 * sampling_rate)
v300ms = int(0.300 * sampling_rate)
v350ms = int(0.350 * sampling_rate)
v200ms = int(0.2 * sampling_rate)
v1200ms = int(1.2 * sampling_rate)
M_th = 0.4 # paper is 0.6
b = np.ones(int(0.02 * sampling_rate)) / 50.
a = [1]
X = ss.filtfilt(b, a, signal)
# 2. Moving averaging of samples in 28 ms interval for electromyogram
# noise suppression a filter with first zero at about 35 Hz.
b = np.ones(int(sampling_rate / 35.)) / 35.
X = ss.filtfilt(b, a, X)
X, _, _ = st.filter_signal(signal=X,
ftype='butter',
band='lowpass',
order=7,
frequency=40.,
sampling_rate=sampling_rate)
X, _, _ = st.filter_signal(signal=X,
ftype='butter',
band='highpass',
order=7,
frequency=9.,
sampling_rate=sampling_rate)
k, Y, L = 1, [], len(X)
for n in range(k + 1, L - k):
Y.append(X[n]**2 - X[n - k] * X[n + k])
Y = np.array(Y)
Y[Y < 0] = 0
b = np.ones(int(sampling_rate / 25.)) / 25.
Y = ss.lfilter(b, a, Y)
# Init
MM = M_th * np.max(Y[:int(5 * sampling_rate)]) * np.ones(5)
MMidx = 0
M = np.mean(MM)
slope = np.linspace(1.0, 0.6, int(sampling_rate))
Rdec = 0
R = 0
RR = np.zeros(5)
RRidx = 0
Rm = 0
QRS = []
Rpeak = []
current_sample = 0
skip = False
F = np.mean(Y[:v350ms])
# Go through each sample
while current_sample < len(Y):
if QRS:
# No detection is allowed 200 ms after the current one. In
# the interval QRS to QRS+200ms a new value of M5 is calculated: newM5 = 0.6*max(Yi)
if current_sample <= QRS[-1] + v200ms:
Mnew = M_th * max(Y[QRS[-1]:QRS[-1] + v200ms])
Mnew = Mnew if Mnew <= 1.5 * MM[MMidx -
1] else 1.1 * MM[MMidx - 1]
MM[MMidx] = Mnew
MMidx = np.mod(MMidx + 1, 5)
# M is calculated as an average value of MM.
Mtemp = np.mean(MM)
M = Mtemp
skip = True
elif current_sample >= QRS[-1] + v200ms and current_sample < QRS[
-1] + v1200ms:
M = Mtemp * slope[current_sample - QRS[-1] - v200ms]
# After 1200 ms M remains unchanged.
# R = 0 V in the interval from the last detected QRS to 2/3 of the expected Rm.
if current_sample >= QRS[-1] and current_sample < QRS[-1] + (
2 / 3.) * Rm:
R = 0
elif current_sample >= QRS[-1] + (
2 / 3.) * Rm and current_sample < QRS[-1] + Rm:
R += Rdec
# After QRS + Rm the decrease of R is stopped
# MFR = M + F + R
MFR = M + F + R
# QRS or beat complex is detected if Yi = MFR
if not skip and Y[current_sample] >= MFR:
QRS += [current_sample]
Rpeak += [QRS[-1] + np.argmax(Y[QRS[-1]:QRS[-1] + v300ms])]
if len(QRS) >= 2:
# A buffer with the 5 last RR intervals is updated at any new QRS detection.
RR[RRidx] = QRS[-1] - QRS[-2]
RRidx = np.mod(RRidx + 1, 5)
skip = False
# With every signal sample, F is updated adding the maximum
# of Y in the latest 50 ms of the 350 ms interval and
# subtracting maxY in the earliest 50 ms of the interval.
if current_sample >= v350ms:
Y_latest50 = Y[current_sample - v50ms:current_sample]
Y_earliest50 = Y[current_sample - v350ms:current_sample - v300ms]
F += (max(Y_latest50) - max(Y_earliest50)) / 1000.
# Rm is the mean value of the buffer RR.
Rm = np.mean(RR)
current_sample += 1
rpeaks = []
for i in Rpeak:
a, b = i - v100ms, i + v100ms
if a < 0:
a = 0
if b > length:
b = length
rpeaks.append(np.argmax(signal[a:b]) + a)
rpeaks = sorted(list(set(rpeaks)))
rpeaks = np.array(rpeaks, dtype='int')
return utils.ReturnTuple((rpeaks, ), ('rpeaks', ))
def smoother(signal=None, kernel='boxzen', size=10, mirror=True, **kwargs):
"""Smooth a signal using an N-point moving average [MAvg]_ filter.
This implementation uses the convolution of a filter kernel with the input
signal to compute the smoothed signal [Smit97]_.
Availabel kernels: median, boxzen, boxcar, triang, blackman, hamming, hann,
bartlett, flattop, parzen, bohman, blackmanharris, nuttall, barthann,
kaiser (needs beta), gaussian (needs std), general_gaussian (needs power,
width), slepian (needs width), chebwin (needs attenuation).
Parameters
----------
signal : array
Signal to smooth.
kernel : str, array, optional
Type of kernel to use; if array, use directly as the kernel.
size : int, optional
Size of the kernel; ignored if kernel is an array.
mirror : bool, optional
If True, signal edges are extended to avoid boundary effects.
``**kwargs`` : dict, optional
Additional keyword arguments are passed to the underlying
scipy.signal.windows function.
Returns
-------
signal : array
Smoothed signal.
params : dict
Smoother parameters.
Notes
-----
* When the kernel is 'median', mirror is ignored.
References
----------
.. [MAvg] Wikipedia, "Moving Average",
http://en.wikipedia.org/wiki/Moving_average
.. [Smit97] S. W. Smith, "Moving Average Filters - Implementation by
Convolution", http://www.dspguide.com/ch15/1.htm, 1997
"""
# check inputs
if signal is None:
raise TypeError("Please specify a signal to smooth.")
length = len(signal)
if isinstance(kernel, six.string_types):
# check length
if size > length:
size = length - 1
if size < 1:
size = 1
if kernel == 'boxzen':
# hybrid method
# 1st pass - boxcar kernel
aux, _ = smoother(signal,
kernel='boxcar',
size=size,
mirror=mirror)
# 2nd pass - parzen kernel
smoothed, _ = smoother(aux,
kernel='parzen',
size=size,
mirror=mirror)
params = {'kernel': kernel, 'size': size, 'mirror': mirror}
args = (smoothed, params)
names = ('signal', 'params')
return utils.ReturnTuple(args, names)
elif kernel == 'median':
# median filter
if size % 2 == 0:
raise ValueError(
"When the kernel is 'median', size must be odd.")
smoothed = ss.medfilt(signal, kernel_size=size)
params = {'kernel': kernel, 'size': size, 'mirror': mirror}
args = (smoothed, params)
names = ('signal', 'params')
return utils.ReturnTuple(args, names)
else:
win = _get_window(kernel, size, **kwargs)
elif isinstance(kernel, np.ndarray):
win = kernel
size = len(win)
# check length
if size > length:
raise ValueError("Kernel size is bigger than signal length.")
if size < 1:
raise ValueError("Kernel size is smaller than 1.")
else:
raise TypeError("Unknown kernel type.")
# convolve
w = win / win.sum()
if mirror:
aux = np.concatenate(
(signal[0] * np.ones(size), signal, signal[-1] * np.ones(size)))
smoothed = np.convolve(w, aux, mode='same')
smoothed = smoothed[size:-size]
else:
smoothed = np.convolve(w, signal, mode='same')
# output
params = {'kernel': kernel, 'size': size, 'mirror': mirror}
params.update(kwargs)
args = (smoothed, params)
names = ('signal', 'params')
return utils.ReturnTuple(args, names)
def compare_segmentation(reference=None,
test=None,
sampling_rate=1000.,
offset=0,
minRR=None,
tol=0.05):
# check inputs
if reference is None:
raise TypeError("Please specify an input reference list of R-peak \
locations.")
if test is None:
raise TypeError("Please specify an input test list of R-peak \
locations.")
if minRR is None:
minRR = np.inf
sampling_rate = float(sampling_rate)
# ensure numpy
reference = np.array(reference)
test = np.array(test)
# convert to samples
minRR = minRR * sampling_rate
tol = tol * sampling_rate
TP = 0
FP = 0
matchIdx = []
dev = []
for i, r in enumerate(test):
# deviation to closest R in reference
ref = reference[np.argmin(np.abs(reference - (r + offset)))]
error = np.abs(ref - (r + offset))
if error < tol:
TP += 1
matchIdx.append(i)
dev.append(error)
else:
if len(matchIdx) > 0:
bdf = r - test[matchIdx[-1]]
if bdf < minRR:
# false positive, but removable with RR interval check
pass
else:
FP += 1
else:
FP += 1
# convert deviations to time
dev = np.array(dev, dtype='float')
dev /= sampling_rate
nd = len(dev)
if nd == 0:
mdev = np.nan
sdev = np.nan
elif nd == 1:
mdev = np.mean(dev)
sdev = 0.
else:
mdev = np.mean(dev)
sdev = np.std(dev, ddof=1)
# interbeat interval
th1 = 1.5 # 40 bpm
th2 = 0.3 # 200 bpm
rIBI = np.diff(reference)
rIBI = np.array(rIBI, dtype='float')
rIBI /= sampling_rate
good = np.nonzero((rIBI < th1) & (rIBI > th2))[0]
rIBI = rIBI[good]
nr = len(rIBI)
if nr == 0:
rIBIm = np.nan
rIBIs = np.nan
elif nr == 1:
rIBIm = np.mean(rIBI)
rIBIs = 0.
else:
rIBIm = np.mean(rIBI)
rIBIs = np.std(rIBI, ddof=1)
tIBI = np.diff(test[matchIdx])
tIBI = np.array(tIBI, dtype='float')
tIBI /= sampling_rate
good = np.nonzero((tIBI < th1) & (tIBI > th2))[0]
tIBI = tIBI[good]
nt = len(tIBI)
if nt == 0:
tIBIm = np.nan
tIBIs = np.nan
elif nt == 1:
tIBIm = np.mean(tIBI)
tIBIs = 0.
else:
tIBIm = np.mean(tIBI)
tIBIs = np.std(tIBI, ddof=1)
# output
perf = float(TP) / len(reference)
acc = float(TP) / (TP + FP)
err = float(FP) / (TP + FP)
args = (TP, FP, perf, acc, err, matchIdx, dev, mdev, sdev, rIBIm, rIBIs,
tIBIm, tIBIs)
names = (
'TP',
'FP',
'performance',
'acc',
'err',
'match',
'deviation',
'mean_deviation',
'std_deviation',
'mean_ref_ibi',
'std_ref_ibi',
'mean_test_ibi',
'std_test_ibi',
)
return utils.ReturnTuple(args, names)
def signal_stats(signal=None):
"""Compute various metrics describing the signal.
Parameters
----------
signal : array
Input signal.
Returns
-------
mean : float
Mean of the signal.
median : float
Median of the signal.
max : float
Maximum signal amplitude.
var : float
Signal variance (unbiased).
std_dev : float
Standard signal deviation (unbiased).
abs_dev : float
Absolute signal deviation.
kurtosis : float
Signal kurtosis (unbiased).
skew : float
Signal skewness (unbiased).
"""
# check inputs
if signal is None:
raise TypeError("Please specify an input signal.")
# ensure numpy
signal = np.array(signal)
# mean
mean = np.mean(signal)
# median
median = np.median(signal)
# maximum amplitude
maxAmp = np.abs(signal - mean).max()
# variance
sigma2 = signal.var(ddof=1)
# standard deviation
sigma = signal.std(ddof=1)
# absolute deviation
ad = np.sum(np.abs(signal - median))
# kurtosis
kurt = stats.kurtosis(signal, bias=False)
# skweness
skew = stats.skew(signal, bias=False)
# output
args = (mean, median, maxAmp, sigma2, sigma, ad, kurt, skew)
names = ('mean', 'median', 'max', 'var', 'std_dev', 'abs_dev', 'kurtosis',
'skewness')
return utils.ReturnTuple(args, names)
def windower(signal=None,
size=None,
step=None,
fcn=None,
fcn_kwargs=None,
kernel='boxcar',
kernel_kwargs=None):
"""Apply a function to a signal in sequential windows, with optional overlap.
Availabel window kernels: boxcar, triang, blackman, hamming, hann,
bartlett, flattop, parzen, bohman, blackmanharris, nuttall, barthann,
kaiser (needs beta), gaussian (needs std), general_gaussian (needs power,
width), slepian (needs width), chebwin (needs attenuation).
Parameters
----------
signal : array
Input signal.
size : int
Size of the signal window.
step : int, optional
Size of window shift; if None, there is no overlap.
fcn : callable
Function to apply to each window.
fcn_kwargs : dict, optional
Additional keyword arguments to pass to 'fcn'.
kernel : str, array, optional
Type of kernel to use; if array, use directly as the kernel.
kernel_kwargs : dict, optional
Additional keyword arguments to pass on window creation; ignored if
'kernel' is an array.
Returns
-------
index : array
Indices characterizing window locations (start of the window).
values : array
Concatenated output of calling 'fcn' on each window.
"""
# check inputs
if signal is None:
raise TypeError("Please specify an input signal.")
if fcn is None:
raise TypeError("Please specify a function to apply to each window.")
if kernel_kwargs is None:
kernel_kwargs = {}
length = len(signal)
if isinstance(kernel, basestring):
# check size
if size > length:
raise ValueError("Window size must be smaller than signal length.")
win = _get_window(kernel, size, **kernel_kwargs)
elif isinstance(kernel, np.ndarray):
win = kernel
size = len(win)
# check size
if size > length:
raise ValueError("Window size must be smaller than signal length.")
if step is None:
step = size
if step <= 0:
raise ValueError("Step size must be at least 1.")
# number of windows
nb = 1 + (length - size) / step
# check signal dimensionality
if np.ndim(signal) == 2:
# time along 1st dim, tile window
nch = np.shape(signal)[1]
win = np.tile(np.reshape(win, (size, 1)), nch)
index = []
values = []
for i in xrange(nb):
start = i * step
stop = start + size
index.append(start)
aux = signal[start:stop] * win
# apply function
out = fcn(aux, **fcn_kwargs)
values.append(out)
# transform to numpy
index = np.array(index, dtype='int')
values = np.array(values)
return utils.ReturnTuple((index, values), ('index', 'values'))
def power_spectrum(signal=None,
sampling_rate=1000.,
pad=None,
pow2=False,
decibel=True):
"""Compute the power spectrum of a signal (one-sided).
Parameters
----------
signal : array
Input signal.
sampling_rate : int, float, optional
Sampling frequency (Hz).
pad : int, optional
Padding for the Fourier Transform (number of zeros added).
pow2 : bool, optional
If True, rounds the number of points `N = len(signal) + pad` to the
nearest power of 2 greater than N.
decibel : bool, optional
If True, returns the power in decibels.
Returns
-------
freqs : array
Array of frequencies (Hz) at which the power was computed.
power : array
Power spectrum.
"""
# check inputs
if signal is None:
raise TypeError("Please specify an input signal.")
npoints = len(signal)
if pad is not None:
if pad >= 0:
npoints += pad
else:
raise ValueError("Padding must be a positive integer.")
# power of 2
if pow2:
npoints = 2**(np.ceil(np.log2(npoints)))
Nyq = float(sampling_rate) / 2
hpoints = npoints / 2
freqs = np.linspace(0, Nyq, hpoints)
power = np.abs(np.fft.fft(signal, npoints)) / npoints
# one-sided
power = power[:hpoints]
power[1:] *= 2
power = np.power(power, 2)
if decibel:
power = 10. * np.log10(power)
return utils.ReturnTuple((freqs, power), ('freqs', 'power'))
def filter_signal(signal=None,
ftype='FIR',
band='lowpass',
order=None,
frequency=None,
sampling_rate=1000.,
**kwargs):
"""Filter a signal according to the given parameters.
Parameters
----------
signal : array
Signal to filter.
ftype : str
Filter type:
* Finite Impulse Response filter ('FIR');
* Butterworth filter ('butter');
* Chebyshev filters ('cheby1', 'cheby2');
* Elliptic filter ('ellip');
* Bessel filter ('bessel').
band : str
Band type:
* Low-pass filter ('lowpass');
* High-pass filter ('highpass');
* Band-pass filter ('bandpass');
* Band-stop filter ('bandstop').
order : int
Order of the filter.
frequency : int, float, list, array
Cutoff frequencies; format depends on type of band:
* 'lowpass' or 'bandpass': single frequency;
* 'bandpass' or 'bandstop': pair of frequencies.
sampling_rate : int, float, optional
Sampling frequency (Hz).
``**kwargs`` : dict, optional
Additional keyword arguments are passed to the underlying
scipy.signal function.
Returns
-------
signal : array
Filtered signal.
sampling_rate : float
Sampling frequency (Hz).
params : dict
Filter parameters.
Notes
-----
* Uses a forward-backward filter implementation. Therefore, the combined
filter has linear phase.
"""
# check inputs
if signal is None:
raise TypeError("Please specify a signal to filter.")
# get filter
b, a = get_filter(ftype=ftype,
order=order,
frequency=frequency,
sampling_rate=sampling_rate,
band=band,
**kwargs)
# filter
filtered, _ = _filter_signal(b, a, signal, check_phase=True)
# output
params = {
'ftype': ftype,
'order': order,
'frequency': frequency,
'band': band,
}
params.update(kwargs)
args = (filtered, sampling_rate, params)
names = ('signal', 'sampling_rate', 'params')
return utils.ReturnTuple(args, names)
def get_filter(ftype='FIR',
band='lowpass',
order=None,
frequency=None,
sampling_rate=1000.,
**kwargs):
"""Compute digital (FIR or IIR) filter coefficients with the given
parameters.
Parameters
----------
ftype : str
Filter type:
* Finite Impulse Response filter ('FIR');
* Butterworth filter ('butter');
* Chebyshev filters ('cheby1', 'cheby2');
* Elliptic filter ('ellip');
* Bessel filter ('bessel').
band : str
Band type:
* Low-pass filter ('lowpass');
* High-pass filter ('highpass');
* Band-pass filter ('bandpass');
* Band-stop filter ('bandstop').
order : int
Order of the filter.
frequency : int, float, list, array
Cutoff frequencies; format depends on type of band:
* 'lowpass' or 'bandpass': single frequency;
* 'bandpass' or 'bandstop': pair of frequencies.
sampling_rate : int, float, optional
Sampling frequency (Hz).
``**kwargs`` : dict, optional
Additional keyword arguments are passed to the underlying
scipy.signal function.
Returns
-------
b : array
Numerator coefficients.
a : array
Denominator coefficients.
See Also:
scipy.signal
"""
# check inputs
if order is None:
raise TypeError("Please specify the filter order.")
if frequency is None:
raise TypeError("Please specify the cutoff frequency.")
if band not in ['lowpass', 'highpass', 'bandpass', 'bandstop']:
raise ValueError(
"Unknown filter type '%r'; choose 'lowpass', 'highpass', \
'bandpass', or 'bandstop'." % band)
# convert frequencies
frequency = _norm_freq(frequency, sampling_rate)
# get coeffs
b, a = [], []
if ftype == 'FIR':
# FIR filter
if order % 2 == 0:
order += 1
a = np.array([1])
if band in ['lowpass', 'bandstop']:
b = ss.firwin(numtaps=order,
cutoff=frequency,
pass_zero=True,
**kwargs)
elif band in ['highpass', 'bandpass']:
b = ss.firwin(numtaps=order,
cutoff=frequency,
pass_zero=False,
**kwargs)
elif ftype == 'butter':
# Butterworth filter
b, a = ss.butter(N=order,
Wn=frequency,
btype=band,
analog=False,
output='ba',
**kwargs)
elif ftype == 'cheby1':
# Chebyshev type I filter
b, a = ss.cheby1(N=order,
Wn=frequency,
btype=band,
analog=False,
output='ba',
**kwargs)
elif ftype == 'cheby2':
# chevyshev type II filter
b, a = ss.cheby2(N=order,
Wn=frequency,
btype=band,
analog=False,
output='ba',
**kwargs)
elif ftype == 'ellip':
# Elliptic filter
b, a = ss.ellip(N=order,
Wn=frequency,
btype=band,
analog=False,
output='ba',
**kwargs)
elif ftype == 'bessel':
# Bessel filter
b, a = ss.bessel(N=order,
Wn=frequency,
btype=band,
analog=False,
output='ba',
**kwargs)
return utils.ReturnTuple((b, a), ('b', 'a'))
def find_intersection(x1=None,
y1=None,
x2=None,
y2=None,
alpha=1.5,
xtol=1e-6,
ytol=1e-6):
"""Find the intersection points between two lines using piecewise
polynomial interpolation.
Parameters
----------
x1 : array
Array of x-coordinates of the first line.
y1 : array
Array of y-coordinates of the first line.
x2 : array
Array of x-coordinates of the second line.
y2 : array
Array of y-coordinates of the second line.
alpha : float, optional
Resolution factor for the x-axis; fraction of total number of
x-coordinates.
xtol : float, optional
Tolerance for the x-axis.
ytol : float, optional
Tolerance for the y-axis.
Returns
-------
roots : array
Array of x-coordinates of found intersection points.
values : array
Array of y-coordinates of found intersection points.
Notes
-----
* If no intersection is found, returns the closest point.
"""
# check inputs
if x1 is None:
raise TypeError("Please specify the x-coordinates of the first line.")
if y1 is None:
raise TypeError("Please specify the y-coordinates of the first line.")
if x2 is None:
raise TypeError("Please specify the x-coordinates of the second line.")
if y2 is None:
raise TypeError("Please specify the y-coordinates of the second line.")
# ensure numpy
x1 = np.array(x1)
y1 = np.array(y1)
x2 = np.array(x2)
y2 = np.array(y2)
if x1.shape != y1.shape:
raise ValueError(
"Input coordinates for the first line must have the same shape.")
if x2.shape != y2.shape:
raise ValueError(
"Input coordinates for the second line must have the same shape.")
# interpolate
p1 = interpolate.BPoly.from_derivatives(x1, y1[:, np.newaxis])
p2 = interpolate.BPoly.from_derivatives(x2, y2[:, np.newaxis])
# combine x intervals
x = np.r_[x1, x2]
x_min = x.min()
x_max = x.max()
npoints = int(len(np.unique(x)) * alpha)
x = np.linspace(x_min, x_max, npoints)
# initial estimates
pd = p1(x) - p2(x)
zerocs, = zero_cross(pd)
pd_abs = np.abs(pd)
zeros = np.nonzero(pd_abs < ytol)[0]
ind = np.unique(np.concatenate((zerocs, zeros)))
xi = x[ind]
# search for solutions
roots = set()
for v in xi:
root, _, ier, _ = optimize.fsolve(_pdiff,
v, (p1, p2),
full_output=True,
xtol=xtol)
if ier == 1 and x_min <= root <= x_max:
roots.add(root[0])
if len(roots) == 0:
# no solution was found => give the best from the initial estimates
aux = np.abs(pd)
bux = aux.min() * np.ones(npoints, dtype='float')
roots, _ = find_intersection(x,
aux,
x,
bux,
alpha=1.,
xtol=xtol,
ytol=ytol)
# compute values
roots = list(roots)
roots.sort()
roots = np.array(roots)
values = np.mean(np.vstack((p1(roots), p2(roots))), axis=0)
return utils.ReturnTuple((roots, values), ('roots', 'values'))
def synchronize(signal1=None, signal2=None):
"""Align two signals based on cross-correlation.
Parameters
----------
signal1 : array
First input signal.
signal2 : array
Second input signal.
Returns
-------
delay : int
Delay (number of samples) of 'signal1' in relation to 'signal2';
if 'delay' < 0 , 'signal1' is ahead in relation to 'signal2';
if 'delay' > 0 , 'signal1' is delayed in relation to 'signal2'.
corr : float
Value of maximum correlation.
synch1 : array
Biggest possible portion of 'signal1' in synchronization.
synch2 : array
Biggest possible portion of 'signal2' in synchronization.
"""
# check inputs
if signal1 is None:
raise TypeError("Please specify the first input signal.")
if signal2 is None:
raise TypeError("Please specify the second input signal.")
n1 = len(signal1)
n2 = len(signal2)
# correlate
corr = np.correlate(signal1, signal2, mode='full')
x = np.arange(-n2 + 1, n1, dtype='int')
ind = np.argmax(corr)
delay = x[ind]
maxCorr = corr[ind]
# get synchronization overlap
if delay < 0:
c = min([n1, len(signal2[-delay:])])
synch1 = signal1[:c]
synch2 = signal2[-delay:-delay + c]
elif delay > 0:
c = min([n2, len(signal1[delay:])])
synch1 = signal1[delay:delay + c]
synch2 = signal2[:c]
else:
c = min([n1, n2])
synch1 = signal1[:c]
synch2 = signal2[:c]
# output
args = (delay, maxCorr, synch1, synch2)
names = ('delay', 'corr', 'synch1', 'synch2')
return utils.ReturnTuple(args, names)
