def DTW_features(pcg,
transitions,
interval='RR',
constraint='sakoe_chiba',
k=0.1,
norm="resample",
pre=None,
downsample_rate=2,
suf=None,
sigma=None):
k = float(k)
if sigma is not None:
sigma = float(sigma)
pcg = dtw_preprocess(pcg, pre=pre)
resize = ALENGTH[interval] if norm == 'resample' else None
intervals = get_intervals(pcg,
transitions,
interval=interval,
resize=resize)
intervals = [resample(i, len(i) // downsample_rate) for i in intervals]
if norm not in ['path', 'resample']:
raise ValueError("Invalid normalization {0}".format(norm))
path_norm = norm == 'path'
dist_matrix = dtw_distances(intervals,
n_jobs=-1,
constraint=constraint,
k=k,
path_norm=path_norm,
normalize=True)
dist_matrix = finite_matrix(dist_matrix)
medoid_index = np.argmin(np.sum(dist_matrix, axis=0))
# Remove the infinite distances
if sigma:
dist_matrix = -affinity(dist_matrix, sigma)
medoid_distances = dist_matrix[:, medoid_index]
medoid_distances = medoid_distances[np.isfinite(medoid_distances)]
m_MDTW, s_MDTW = _mean_std(medoid_distances)
Q1_MDTW, Q2_MDTW, Q3_MDTW = np.percentile(medoid_distances, [25, 50, 75])
contiguous_distances = np.array(
[dist_matrix[i, i + 1] for i in np.arange(len(dist_matrix) - 1)])
contiguous_distances = contiguous_distances[np.isfinite(
contiguous_distances)]
Q1_CDTW, Q2_CDTW, Q3_CDTW = np.percentile(contiguous_distances,
[25, 50, 75])
features = np.array(
[m_MDTW, s_MDTW, Q1_MDTW, Q2_MDTW, Q3_MDTW, Q1_CDTW, Q2_CDTW, Q3_CDTW])
return features
def random_choice(pcgs, states, records, size):
t_records = list(np.random.choice(records, size, replace=False))
indexes = []
for r in t_records:
pcg = pcgs[r]
transitions = get_transitions(states[r])
intervals = get_intervals(pcg, transitions)
i = np.random.choice(range(len(intervals)))
indexes.append(i)
return t_records, indexes
def cinter_DTW_features(pcg,
transitions,
templates='random_templates',
pre=None,
constraint='sakoe_chiba',
k=0.1,
suf=None,
sigma=None):
k = float(k)
if sigma is not None:
sigma = float(sigma)
templates = custom_loadmat(TEMPLATE_FOLDER + templates)
inter_DTW_features.names = [
"%s_d%02d" % (q, i) for q in ['mean', 'std']
for i, _ in enumerate(templates)
]
pcg = dtw_preprocess(pcg, pre=pre)
distances = []
for interval in ['RR', 'S1', 'Sys', 'S2', 'Dia']:
templates_i = templates[interval]
intervals = get_intervals(pcg,
transitions,
interval=interval,
resize=ALENGTH[interval] // interDTW_down)
intervals = [
resample(i, ALENGTH[interval] // interDTW_down) for i in intervals
][:50]
dist_matrix = dtw_distances(intervals,
templates_i,
n_jobs=-1,
constraint=constraint,
k=k)
dist_matrix = finite_matrix(dist_matrix)
if sigma:
dist_matrix = -affinity(dist_matrix, sigma)
distances.append(dist_matrix)
RR_mean, S1_mean, Sys_mean, S2_mean, Dia_mean = [
np.mean(d) for d in distances
]
RR_std, S1_std, Sys_std, S2_std, Dia_std = [np.std(d) for d in distances]
features = [
RR_mean, S1_mean, Sys_mean, S2_mean, Dia_mean, RR_std, S1_std, Sys_std,
S2_std, Dia_std
]
return features
def build_templates(pcgs, states, records, indexes):
templates = defaultdict(list)
for r, i in tqdm(zip(records, indexes), total=len(records)):
pcg = pcgs[r]
for interval in ['RR', 'S1', 'Sys', 'S2', 'Dia']:
transitions = get_transitions(states[r])
inter = get_intervals(pcg,
transitions,
interval=interval,
resize=ALENGTH[interval] // interDTW_down)[i]
templates[interval].append(inter)
return dict(templates)
def mfcc_coefs(pcg, transitions, interval='RR'):
fs = 1000
pcg = _normalize(pcg)
intervals = get_intervals(pcg,
transitions,
interval=interval,
resize=ALENGTH[interval])
signal = np.concatenate(intervals)
win = ALENGTH[interval] / fs
mel = mfcc(signal,
samplerate=fs,
winlen=win,
winstep=win,
numcep=13,
nfilt=26,
nfft=1024)
return mel
def inter_DTW_features(pcg,
transitions,
interval='RR',
templates='random_templates.mat',
pre=None,
constraint='sakoe_chiba',
k=0.1,
suf=None,
sigma=None):
k = float(k)
if sigma is not None:
sigma = float(sigma)
templates = custom_loadmat(TEMPLATE_FOLDER + templates)[interval]
inter_DTW_features.names = [
"%s_d%02d" % (q, i) for q in ['mean', 'std']
for i, _ in enumerate(templates)
]
pcg = dtw_preprocess(pcg, pre=pre)
intervals = get_intervals(pcg,
transitions,
interval=interval,
resize=ALENGTH[interval] // interDTW_down)
dist_matrix = dtw_distances(intervals,
templates,
n_jobs=-1,
constraint=constraint,
k=k)
dist_matrix = finite_matrix(dist_matrix)
if sigma:
dist_matrix = -affinity(dist_matrix, sigma)
mean_dtw = np.mean(dist_matrix, axis=0)
std_dtw = np.mean(dist_matrix, axis=0)
features = np.concatenate([mean_dtw, std_dtw])
return features
def wavelet_features(pcg,
transitions,
interval='RR',
wavelet='rbio3.9',
level=3,
suf=None):
level = int(level)
pcg = _normalize(pcg)
intervals = get_intervals(pcg, transitions, interval=interval)
cAs, cD1s, cD2s, cD3s = [], [], [], []
for x in intervals:
x = np.concatenate((x, np.zeros(2**level - len(x) % 2**level)))
cA, cD1, cD2, cD3 = pywt.wavedec(x, wavelet, level=level)
cAs.append(np.linalg.norm(cA)**2)
cD1s.append(np.linalg.norm(cD1)**2)
cD2s.append(np.linalg.norm(cD2)**2)
cD3s.append(np.linalg.norm(cD3)**2)
m_cA, std_cA = _mean_std(cAs)
m_cD1, std_cD1 = _mean_std(cD1s)
m_cD2, std_cD2 = _mean_std(cD2s)
m_cD3, std_cD3 = _mean_std(cD3s)
return [m_cA, std_cA, m_cD1, std_cD1, m_cD2, std_cD2, m_cD3, std_cD3]
pre = 'env'
fpcgs = {r: dtw_preprocess(pcgs[r], pre=pre) for r in tqdm(records)}
if arguments['-o']:
filename = arguments['-o']
savemat(filename, fpcgs)
print('Saved to {0}'.format(filename))
elif arguments['medoids']:
filename = filename.replace('PCG', 'medoids')
inters = {
r: get_intervals(pcgs[r],
get_transitions(states[r]),
interval='RR',
resize=ALENGTH['RR'] // interDTW_down)
for r in tqdm(records)
}
medoid_indexes = [
dtw_medoid(inters[r], **dtw_params) for r in tqdm(records)
]
medoids = {r: inters[r][m] for r, m in zip(records, medoid_indexes)}
if arguments['-o']:
filename = arguments['-o']
savemat(filename, medoids)
print('Saved to {0}'.format(filename))