def load(self, backup_path=None, num_records=None, checkpoint_interval=100):
samples_per_second = self.samples_per_second
total_sample_size_seconds = self.sample_size_seconds * self.seq_len
to_wavelet = self.to_wavelet
if backup_path is not None:
filename = ('{}.pt' if self.to_wavelet is None else '{}_transformed.pt').format(self.dataset_name)
file_path = os.path.join(backup_path, filename)
if not os.path.exists(backup_path):
os.makedirs(backup_path)
else:
file_path = None
if file_path is not None and os.path.isfile(file_path):
data = torch.load(file_path)
print('Loaded from backup')
self.samples, self.labels = data['samples'], data['labels']
return
print('Reading records. sample size={}, frequency={}'.format(total_sample_size_seconds, samples_per_second))
data, labels = read_records(self.dataset_name, self.data_path,
sample_size_seconds=total_sample_size_seconds,
samples_per_second=samples_per_second, num_records=num_records)
labels = torch.tensor(labels)
print(labels.shape)
print(len(data))
data = [split_sample(sample, self.sample_size_seconds) for sample in data]
count = len(data)
transformed_data = []
print('Preparing {} samples'.format(count))
num_samples = 0
for sample_idx, sample in tqdm(enumerate(data[:count]), desc='Preprocessing examples'):
wavelets = []
for signal_idx, signal in enumerate(sample):
signal = wfdb_processing.normalize_bound(signal.numpy())
if to_wavelet is not None:
sw = to_wavelet(signal)
else:
sw = signal
wavelets.append(torch.tensor(sw))
t = torch.stack(wavelets)
transformed_data.append(t)
num_samples += 1
transformed_data = torch.stack(transformed_data)
if backup_path is not None:
torch.save({
'samples': transformed_data,
'labels': labels
}, file_path)
self.samples, self.labels = transformed_data, labels
def test_normalize_bound(self):
sig, _ = wfdb.rdsamp('sample-data/100')
lb = -5
ub = 15
x = processing.normalize_bound(sig[:, 0], lb, ub)
assert x.shape[0] == sig.shape[0]
assert np.min(x) >= lb
assert np.max(x) <= ub
def test_normalize_bound(self):
sig, _ = wfdb.rdsamp('sample-data/100')
lb = -5
ub = 15
x = processing.normalize_bound(sig[:, 0], lb, ub)
assert x.shape[0] == sig.shape[0]
assert np.min(x) >= lb
assert np.max(x) <= ub
def tsaug_generator(X_all, y_all, batch_size):
"""
Generate Time Series data with sequence labels
Data generator that yields training data as batches.
1. Randomly selects one sample from time series signals
2. Applies time series augmentations to X and Y
3. Normalizes result for X data
4. Reshapes data into correct format for training
Parameters
----------
X_all : 3D numpy array
(N, seqlen, features=1)
y_all : 3D numpy array (binary labels for my case)
(N, seqlen, classes=1)
batch_size : int
Number of training examples in the batch
Yields
------
(X, y) : tuple
Contains training samples with corresponding labels
"""
while True:
X = []
y = []
while len(X) < batch_size:
random_sig_idx = np.random.randint(0, X_all.shape[0])
x1 = X_all[random_sig_idx].flatten()
y1 = y_all[random_sig_idx].flatten()
# Augment X and y and normalize it again
X_aug, y_aug = my_augmenter.augment(x1, y1)
X_aug = normalize_bound(X_aug, lb=-1, ub=1)
X.append(X_aug)
y.append(y_aug)
X = np.asarray(X)
y = np.asarray(y)
X = X.reshape(X.shape[0], X.shape[1], 1)
y = y.reshape(y.shape[0], y.shape[1], 1).astype(int)
yield (X, y)
def ecg_generator(name, signals, wgn, ma, bw, win_size, batch_size):
"""
Generate ECG data with R-peak labels.
Data generator that yields training data as batches. Every instance
of training batch is composed as follows:
1. Randomly select one ECG signal from given list of ECG signals
2. Randomly select one window of given win_size from selected signal
3. Check that window has at least one beat and that all beats are
labled as normal
4. Create label window corresponding the selected window
-beats and four samples next to beats are labeled as 1 while
rest of the samples are labeled as 0
5. Normalize selected signal window from -1 to 1
6. Add noise into signal window and normalize it again to (-1, 1)
7. Add noisy signal and its labels to trainig batch
8. Transform training batches to arrays of needed shape and yield
training batch with corresponding labels when needed
Parameters
----------
signals : list
List of ECG signals
peaks : list
List of peaks locations for the ECG signals
labels : list
List of labels (peak types) for the peaks
ma : array
Muscle artifact signal
bw : array
Baseline wander signal
win_size : int
Number of time steps in the training window
batch_size : int
Number of training examples in the batch
Yields
------
(X, y) : tuple
Contains training samples with corresponding labels
"""
print('processing')
while True:
x = []
section = []
noise_list = []
while len(x) < batch_size:
random_sig_idx = np.random.randint(0, len(signals))
random_sig = signals[random_sig_idx]
#print(random_sig_idx)
# Select one window
beg = np.random.randint(random_sig.shape[0] - win_size)
end = beg + win_size
#section = normalize_bound(section, lb=-1, ub=1)
section.append(random_sig[beg:end])
# Select data for window and normalize it (-1, 1)
data_win = normalize_bound(random_sig[beg:end], lb=-1, ub=1)
# Add noise into data window and normalize it again
added_noise = get_noise(name, wgn, ma, bw, win_size)
added_noise = normalize_bound(added_noise, lb=-1, ub=1)
noise_list.append(added_noise)
data_win = data_win + added_noise
data_win = normalize_bound(data_win, lb=-1, ub=1)
x.append(data_win)
section = np.asarray(section)
section = section.reshape(section.shape[0], section.shape[1], 1)
#print(section, section.shape, type(section))
x = np.asarray(x)
print(x.shape)
x = x.reshape(x.shape[0], x.shape[1], 1)
#print(x)
id = np.random.randint(0, len(x))
plt.subplot(311)
plt.plot(section[id])
plt.title('original ' + name + ' ecg')
plt.subplot(312)
plt.plot(x[id])
plt.title('noised ' + name + ' ecg')
plt.subplot(313)
plt.title('added noise')
plt.plot(noise_list[id])
#plt.savefig(name+'.png')
plt.show()
print('shape:', x.shape)
serialization(name + '_original_ecg', section)
serialization(name + '_noised_ecg', x)
return x, section
def utama(filename):
global timenow
path_data = filename
path_model = "Best Model 3 Classes.h5"
ext = path_data.split(".")[-1]
namefile = path_data.split('/')[4].split('.')[0]
print(filename, ext, namefile)
if ext == 'xml':
with open(path_data) as fopen:
file = fopen.read()
data = xmltodict.parse(file, encoding='latin-1')
raw = data['CardioXP']['StudyInfo']['WaveData'][1]['Data']
signal = raw.split(' ')
signal = np.array(signal, dtype=int)
plt.figure(figsize=(15, 10))
plt.plot(range(2700), signal[:2700])
saving_file_bef = './app/static/AF/' + namefile + '-' + timenow + '-bef.png'
plt.savefig(saving_file_bef)
signal = normalize_bound(signal, lb=0, ub=1)
signal = denoising(signal)
signal = signal[0:2700]
plt.figure(figsize=(15, 10))
plt.plot(range(2700), signal[:2700])
saving_file_aft = './app/static/AF/' + namefile + '-' + timenow + '-aft.png'
plt.savefig(saving_file_aft)
elif ext == 'dat':
record = wfdb.rdrecord(path_data)
record_dict = record.__dict__
p_signal = record_dict['p_signal'][:, 0]
plt.figure(figsize=(15, 10))
plt.plot(range(2700), p_signal[:2700])
saving_file_bef = './app/static/AF/' + namefile + '-' + timenow + '-bef.png'
plt.savefig(saving_file_bef)
p_signal = normalize_bound(p_signal, lb=0, ub=1)
p_signal = denoising(p_signal)
signal = p_signal[0:2700]
plt.figure(figsize=(15, 10))
plt.plot(range(2700), signal[:2700])
saving_file_aft = './app/static/AF/' + namefile + '-' + timenow + '-aft.png'
plt.savefig(saving_file_aft)
elif ext == 'mat':
data = scipy.io.loadmat(path_data)
sampels = data['val'][0]
plt.figure(figsize=(15, 10))
plt.plot(range(2700), sampels[:2700])
saving_file_bef = './app/static/AF/' + namefile + '-' + timenow + '-bef.png'
plt.savefig(saving_file_bef)
sampels = normalize_bound(sampels, lb=0, ub=1)
sampels = denoising(sampels)
signal = sampels[0:2700]
plt.figure(figsize=(15, 10))
plt.plot(range(2700), signal[:2700])
saving_file_aft = './app/static/AF/' + namefile + '-' + timenow + '-aft.png'
plt.savefig(saving_file_aft)
else:
sys.exit("Not A Valid Data")
if len(signal) < 2700:
sys.exit("Signal length is not sufficient")
signal = np.reshape(signal, (1, 2700, 1))
model = load_model(path_model)
predicted_class = model.predict_classes(signal)[0]
if predicted_class == 0:
clear_session()
print("Normal")
return 'Normal', namefile + '-' + timenow
elif predicted_class == 1:
clear_session()
print("Atrial Fibrilation")
return 'Atrial Fibrilation', namefile + '-' + timenow
else:
clear_session()
print("Non AF")
return 'Non AF', namefile + '-' + timenow
for item in np.arange(len(path_split)):
record = wfdb.rdrecord(path_split[item])
record_dict = record.__dict__
signal = record_dict['p_signal'][:, 0]
annotation = wfdb.rdann(path_split[item], 'atr')
ann_dict = annotation.__dict__
symbol = ann_dict['symbol']
peaks = ann_dict['sample']
name = ann_dict['record_name']
fs = ann_dict['fs']
t1 = np.int(0.25 * fs)
t2 = np.int(0.45 * fs)
peak = np.arange(len(peaks))
new_signal = utility.wavelet_transform(signal, 8, 'sym5')
new_signal = normalize_bound(new_signal)
beats = []
labels = []
for x in peak:
if (peaks[x] - t1) > 0 and (peaks[x] +
t2) < len(signal): #Ini segmentasi
#if (symbol[x] == 'A' or symbol[x] == 'L' or symbol[x] == 'N' or symbol[x] == '.' or symbol[x] == 'P' or symbol[x] == 'R' or symbol[x] == 'V' or symbol[x] == 'f' or symbol[x] == 'F' or symbol[x] == '!' or symbol[x] == 'j' ):
if (symbol[x] == '~' or symbol[x] == '|' or symbol[x] == '+'
or symbol[x] == 'B' or symbol[x] == 'F' or symbol[x] == 'f'
or symbol[x] == 'Q' or symbol[x] == 'a'
or symbol[x] == 'J'):
continue
else:
beat = new_signal[peaks[x] - t1:peaks[x] + t2]
def dataGeneration(data_path, csv_path, record_path):
# initialize dataset
dataset = pd.DataFrame(columns=['label', 'record'])
if record_path == None:
# a loop for each patient
detail_path = data_path + '/'
record_files = [
i.split('.')[0] for i in os.listdir(detail_path)
if (not i.startswith('.') and i.endswith('.hea'))
]
Bar.check_tty = False
bar = Bar('Processing',
max=len(record_files),
fill='#',
suffix='%(percent)d%%')
# a loop for each record
for record_name in record_files:
# load record
signal, info = wfdb.rdsamp(detail_path + record_name)
fs = 200
signal = processing.resample_sig(signal[:, 0], info['fs'], fs)[0]
# set some parameters
window_size_half = int(fs * 0.125 / 2)
max_bpm = 230
# detect QRS peaks
qrs_inds = processing.gqrs_detect(signal, fs=fs)
search_radius = int(fs * 60 / max_bpm)
corrected_qrs_inds = processing.correct_peaks(
signal,
peak_inds=qrs_inds,
search_radius=search_radius,
smooth_window_size=150)
average_qrs = 0
count = 0
for i in range(1, len(corrected_qrs_inds) - 1):
start_ind = corrected_qrs_inds[i] - window_size_half
end_ind = corrected_qrs_inds[i] + window_size_half + 1
if start_ind < corrected_qrs_inds[
i - 1] or end_ind > corrected_qrs_inds[i + 1]:
continue
average_qrs = average_qrs + signal[start_ind:end_ind]
count = count + 1
# remove outliers
if count < 8:
print('\noutlier detected, discard ' + record_name)
continue
average_qrs = average_qrs / count
corrcoefs = []
for i in range(1, len(corrected_qrs_inds) - 1):
start_ind = corrected_qrs_inds[i] - window_size_half
end_ind = corrected_qrs_inds[i] + window_size_half + 1
if start_ind < corrected_qrs_inds[
i - 1] or end_ind > corrected_qrs_inds[i + 1]:
corrcoefs.append(-100)
continue
corrcoef = pearsonr(signal[start_ind:end_ind], average_qrs)[0]
corrcoefs.append(corrcoef)
max_corr = list(map(corrcoefs.index, heapq.nlargest(8, corrcoefs)))
index_corr = random.sample(
list(itertools.permutations(max_corr, 8)), 100)
for index in index_corr:
# a temp dataframe to store one record
record_temp = pd.DataFrame()
signal_temp = []
for i in index:
start_ind = corrected_qrs_inds[i + 1] - window_size_half
end_ind = corrected_qrs_inds[i + 1] + window_size_half + 1
sig = processing.normalize_bound(signal[start_ind:end_ind],
-1, 1)
signal_temp = np.concatenate((signal_temp, sig))
record_temp = record_temp.append(pd.DataFrame(
signal_temp.reshape(-1, signal_temp.shape[0])),
ignore_index=True,
sort=False)
record_temp['label'] = record_name
record_temp['record'] = record_name
# add it to final dataset
dataset = dataset.append(record_temp,
ignore_index=True,
sort=False)
bar.next()
bar.finish()
else:
patient_folders = [
i for i in os.listdir(data_path)
if (not i.startswith('.') and i.startswith(record_path))
]
Bar.check_tty = False
bar = Bar('Processing',
max=len(patient_folders),
fill='#',
suffix='%(percent)d%%')
# a loop for each patient
for patient_name in patient_folders:
detail_path = data_path + patient_name + '/'
record_files = [
i.split('.')[0] for i in os.listdir(detail_path)
if i.endswith('.hea')
]
# a loop for each record
for record_name in record_files:
# load record
signal, info = wfdb.rdsamp(detail_path + record_name)
fs = 200
signal = processing.resample_sig(signal[:, 0], info['fs'],
fs)[0]
# set some parameters
window_size_half = int(fs * 0.125 / 2)
max_bpm = 230
# detect QRS peaks
qrs_inds = processing.gqrs_detect(signal, fs=fs)
search_radius = int(fs * 60 / max_bpm)
corrected_qrs_inds = processing.correct_peaks(
signal,
peak_inds=qrs_inds,
search_radius=search_radius,
smooth_window_size=150)
average_qrs = 0
count = 0
for i in range(1, len(corrected_qrs_inds) - 1):
start_ind = corrected_qrs_inds[i] - window_size_half
end_ind = corrected_qrs_inds[i] + window_size_half + 1
if start_ind < corrected_qrs_inds[
i - 1] or end_ind > corrected_qrs_inds[i + 1]:
continue
average_qrs = average_qrs + signal[start_ind:end_ind]
count = count + 1
# remove outliers
if count < 8:
print('\noutlier detected, discard ' + record_name +
' of ' + patient_name)
continue
average_qrs = average_qrs / count
corrcoefs = []
for i in range(1, len(corrected_qrs_inds) - 1):
start_ind = corrected_qrs_inds[i] - window_size_half
end_ind = corrected_qrs_inds[i] + window_size_half + 1
if start_ind < corrected_qrs_inds[
i - 1] or end_ind > corrected_qrs_inds[i + 1]:
corrcoefs.append(-100)
continue
corrcoef = pearsonr(signal[start_ind:end_ind],
average_qrs)[0]
corrcoefs.append(corrcoef)
max_corr = list(
map(corrcoefs.index, heapq.nlargest(8, corrcoefs)))
index_corr = random.sample(
list(itertools.permutations(max_corr, 8)), 100)
for index in index_corr:
# a temp dataframe to store one record
record_temp = pd.DataFrame()
signal_temp = []
for i in index:
start_ind = corrected_qrs_inds[i +
1] - window_size_half
end_ind = corrected_qrs_inds[i +
1] + window_size_half + 1
sig = processing.normalize_bound(
signal[start_ind:end_ind], -1, 1)
signal_temp = np.concatenate((signal_temp, sig))
record_temp = record_temp.append(pd.DataFrame(
signal_temp.reshape(-1, signal_temp.shape[0])),
ignore_index=True,
sort=False)
record_temp['label'] = patient_name
record_temp['record'] = record_name
# add it to final dataset
dataset = dataset.append(record_temp,
ignore_index=True,
sort=False)
bar.next()
bar.finish()
# save for further use
dataset.to_csv(csv_path, index=False)
print('processing completed')
def ecg_generator(signals, peaks, labels, ma, bw, win_size, batch_size):
"""
Generate ECG data with R-peak labels.
Data generator that yields training data as batches. Every instance
of training batch is composed as follows:
1. Randomly select one ECG signal from given list of ECG signals
2. Randomly select one window of given win_size from selected signal
3. Check that window has at least one beat and that all beats are
labled as normal
4. Create label window corresponding the selected window
-beats and four samples next to beats are labeled as 1 while
rest of the samples are labeled as 0
5. Normalize selected signal window from -1 to 1
6. Add noise into signal window and normalize it again to (-1, 1)
7. Add noisy signal and its labels to trainig batch
8. Transform training batches to arrays of needed shape and yield
training batch with corresponding labels when needed
Parameters
----------
signals : list
List of ECG signals
peaks : list
List of peaks locations for the ECG signals
labels : list
List of labels (peak types) for the peaks
ma : array
Muscle artifact signal
bw : array
Baseline wander signal
win_size : int
Number of time steps in the training window
batch_size : int
Number of training examples in the batch
Yields
------
(X, y) : tuple
Contains training samples with corresponding labels
"""
while True:
X = []
y = []
while len(X) < batch_size:
random_sig_idx = np.random.randint(0, len(signals))
random_sig = signals[random_sig_idx]
p4sig = peaks[random_sig_idx]
plabels = labels[random_sig_idx]
# Select one window
beg = np.random.randint(random_sig.shape[0]-win_size)
end = beg + win_size
# Select peaks that fall into selected window.
# Buffer of 3 to the window edge is needed as labels are
# inserted also next to point)
p_in_win = p4sig[(p4sig >= beg+3) & (p4sig <= end-3)]-beg
# Check that there is at least one peak in the window
if p_in_win.shape[0] >= 1:
# Select labels that fall into selected window
lab_in_win = plabels[(p4sig >= beg+3) & (p4sig <= end-3)]
# Check that every beat in the window is normal beat
if np.all(lab_in_win == 1):
# Create labels for data window
window_labels = np.zeros(win_size)
np.put(window_labels, p_in_win, lab_in_win)
# Put labels also next to peak
np.put(window_labels, p_in_win+1, lab_in_win)
np.put(window_labels, p_in_win+2, lab_in_win)
np.put(window_labels, p_in_win-1, lab_in_win)
np.put(window_labels, p_in_win-2, lab_in_win)
# Select data for window and normalize it (-1, 1)
data_win = normalize_bound(random_sig[beg:end],
lb=-1, ub=1)
# Add noise into data window and normalize it again
data_win = data_win + get_noise(ma, bw, win_size)
data_win = normalize_bound(data_win, lb=-1, ub=1)
X.append(data_win)
y.append(window_labels)
X = np.asarray(X)
y = np.asarray(y)
X = X.reshape(X.shape[0], X.shape[1], 1)
y = y.reshape(y.shape[0], y.shape[1], 1).astype(int)
yield (X, y)
