def new_sample(path, name, sampling_rate, i):
x_signal, y_signal = read_sample(path,
name,
sampling_rate,
preprocess=False)
rpeaks = detect_r_points(y_signal, y_signal.shape[0], sampling_rate)
print('y_shape %d' % y_signal.shape[0])
print('rpeaks_shape %d' % rpeaks.shape[0])
label, type = classify_by_rpoints(y_signal, rpeaks, sampling_rate)
'''
plot_length = 40000
trace1 = go.Scatter(y=y_signal[:plot_length], x=x_signal[:plot_length], name='Signal')
trace2 = go.Scatter(y=y_signal[rpeaks], x=x_signal[rpeaks],mode='markers', name='rpeaks')
figure = go.Figure(data=[trace1, trace2])
py.plot(figure, filename=name)
print('Plotting is done! :)')
'''
print('% s Suggested label by simple handy classification = %s' %
(name, label))
#if label == 'Arrhythmic':
if 1 == 1:
trace1 = go.Scatter(y=y_signal[:], x=x_signal[:], name='Signal')
trace2 = go.Scatter(y=y_signal[rpeaks[:, 0]],
x=x_signal[rpeaks[:, 0]],
mode='markers',
name='rpeaks')
layout = go.Layout(title=name)
figure = go.Figure(data=[trace1, trace2], layout=layout)
py.plot(figure, filename=name)
#print('Plotting is done! :) sampling rate = %s' %sampling_rates[i])
return label
def load_sample_test(path, name, sampling_rate=1000, dimension=5000, step=2):
x_signal, y_signal = read_sample(path, name, sampling_rate)
length = y_signal.shape[0]
rpeaks = detect_r_points(y_signal, length, sampling_rate)
sample_x = np.empty([0, dimension])
for index in range(0, len(rpeaks), step):
if rpeaks[index] + dimension < length:
x = np.empty([dimension])
x = y_signal[rpeaks[index]:rpeaks[index] + dimension, 0]
sample_x = np.append(sample_x, np.array([x]), axis=0)
return sample_x
def load_sample_raw(path,
name,
real_label,
sampling_rate,
explanation,
dimension1=4,
dimension2=500,
step=500,
train=True):
x_signal, y_signal = read_sample(path,
name,
sampling_rate,
preprocess=False)
length = y_signal.shape[0]
sample_x = np.empty([0, dimension1, dimension2])
sample_y = np.empty([0, 1])
if real_label == 'Normal':
label = 0
else:
label = 1
episode = dimension1 * dimension2
if (real_label == 'Normal') | (train is False) | (
explanation == 'Bradicardia') | (explanation == 'Tachicardia') | (
explanation == 'long QT'):
pointer = 10
while pointer + episode < length:
temp = y_signal[pointer:pointer + episode, 0]
x = np.reshape(temp, [dimension1, dimension2])
y = np.array(label)
sample_x = np.append(sample_x, np.array([x]), axis=0)
sample_y = np.append(sample_y, np.array(label))
pointer += step
else:
rpeaks = detect_r_points(y_signal, length, sampling_rate)
problems = analyze_rpoints(y_signal, rpeaks, sampling_rate)
for p in problems:
problem = int(p)
pointer = max(problem - episode + dimension2, 10)
while pointer + episode < min(length, problem + 7000):
x = y_signal[pointer:pointer + episode, 0]
y = np.array(label)
sample_x = np.append(sample_x, np.array([x]), axis=0)
sample_y = np.append(sample_y, np.array(label))
pointer += step
sample_y = np.reshape(sample_y, [sample_y.shape[0], 1])
sample_x = np.reshape(sample_x,
[sample_x.shape[0], dimension1, dimension2])
return sample_x, sample_y
def load_sample(path,
name,
real_label,
sampling_rate,
explanation,
interval,
dimension=4000,
step=2,
train=False,
preprocess=False):
x_signal, y_signal = read_sample(path, name, sampling_rate, preprocess)
length = y_signal.shape[0]
rpeaks = detect_r_points(y_signal, length, sampling_rate)
sample_x = np.empty([0, dimension])
sample_y = np.empty([0, 1])
if real_label == 'Normal':
label = 0
else:
label = 1
for index in range(0, len(rpeaks), step):
if rpeaks[index] + dimension < length:
if (train is False) | (real_label == 'Normal') | (
explanation == 'Bradicardia'
) | (explanation == 'Tachicardia') | (explanation == 'Long QT'):
x = np.empty([dimension])
x = y_signal[rpeaks[index]:rpeaks[index] + dimension, 0]
y = np.array(label)
sample_x = np.append(sample_x, np.array([x]), axis=0)
sample_y = np.append(sample_y, np.array(label))
elif check_overlap(interval, x_signal[rpeaks[index]],
x_signal[rpeaks[index] + dimension]):
x = np.empty([dimension, 1])
x = y_signal[rpeaks[index]:rpeaks[index] + dimension, 0]
y = np.array(label)
sample_x = np.append(sample_x, np.array([x]), axis=0)
sample_y = np.append(sample_y, np.array(label))
return sample_x, sample_y
def check_new_input(file_path, file_name, sampling_rate=1000, plot=True):
#Create and load networks
dimension = 5000
dimension_fraction = 1
cnn_model = Sequential()
cnn_model.add(
Conv1D(filters=16,
kernel_size=80,
strides=2,
padding="same",
input_shape=(dimension // dimension_fraction,
dimension_fraction)))
cnn_model.add(BatchNormalization())
cnn_model.add(Activation('selu'))
cnn_model.add(MaxPooling1D())
cnn_model.add(Dropout(0.5))
cnn_model.add(Conv1D(filters=16, kernel_size=80, padding="same"))
cnn_model.add(BatchNormalization())
cnn_model.add(Activation('selu'))
cnn_model.add(Dropout(0.5))
cnn_model.add(Conv1D(filters=16, kernel_size=80, padding="same"))
cnn_model.add(BatchNormalization())
cnn_model.add(Activation('selu'))
cnn_model.add(MaxPooling1D())
cnn_model.add(Dropout(0.5))
cnn_model.add(Flatten())
cnn_model.add(Dense(1))
cnn_model.add(Activation('sigmoid'))
adam = Adam(lr=0.0005)
cnn_model.compile(loss='binary_crossentropy',
optimizer=adam,
metrics=['binary_accuracy'])
cnn_model.load_weights('./Trained_networks/cnn_model.h5')
rnn_model = Sequential()
init_one = keras.initializers.Ones()
rnn_model.add(GRU(3, input_shape=(None, 1), kernel_initializer=init_one))
rnn_model.add(BatchNormalization())
rnn_model.add(PReLU())
rnn_model.add(Dropout(0.5))
rnn_model.add(Dense(1))
rnn_model.add(Activation('sigmoid'))
rnn_model.compile(loss='binary_crossentropy',
optimizer='adam',
metrics=['binary_accuracy'])
rnn_model.load_weights('./Trained_networks/rnn_model.h5')
#Classify new sample by CNN-RNN method
sample_x = load_sample_test(file_path,
file_name,
sampling_rate=sampling_rate)
if sample_x.shape[0] > 0:
sample_x = np.reshape(sample_x, [
sample_x.shape[0], dimension // dimension_fraction,
dimension_fraction
])
step1_predicted = cnn_model.predict(np.array(sample_x))
reshaped_predicted = np.reshape(
step1_predicted,
[1, step1_predicted.shape[0], step1_predicted.shape[1]])
final_predicted = rnn_model.predict(reshaped_predicted)
if final_predicted < 0.5:
predicted_label = 'Normal'
else:
predicted_label = 'Arrhythmic'
#Plot signal
if plot == True:
x_signal, y_signal = read_sample(file_path,
file_name,
sampling_rate=sampling_rate,
preprocess=False)
trace1 = go.Scatter(y=y_signal[:], x=x_signal[:], name='Signal')
layout = go.Layout(title=file_name)
figure = go.Figure(data=[trace1], layout=layout)
py.plot(figure, filename=file_name)
print('Plotting is done! :) ')
print('Predicted label = %s' % predicted_label)
return predicted_label
def load_sample_rbased(path,
name,
real_label,
sampling_rate,
explanation,
dimension1=4,
dimension2=500,
step=2,
train=False):
x_signal, y_signal = read_sample(path,
name,
sampling_rate,
preprocess=False)
length = y_signal.shape[0]
rpeaks = detect_r_points(y_signal, length, sampling_rate)
counter_label0 = 0
counter_label1 = 0
counter_sample0 = 0
counter_sample1 = 0
sample_x = np.empty([0, dimension1, dimension2])
sample_y = np.empty([0, 1])
if real_label == 'Normal':
label = 0
counter_label0 += 1
if real_label == 'Arrhythmic':
print('oh oh label 0 , %s' % (real_label))
else:
label = 1
counter_label1 += 1
if real_label == 'Normal':
print('oh oh label 1 , %s' % (real_label))
if (real_label == 'Normal') | (train is False) | (
explanation == 'Bradicardia') | (explanation == 'Tachicardia') | (
explanation == 'long QT'):
for index in range(0, len(rpeaks) - dimension1, step):
if rpeaks[index + dimension1] + dimension2 < length:
x = np.empty([dimension1, dimension2])
for i in range(dimension1):
x[i] = y_signal[rpeaks[i + index]:rpeaks[i + index] +
dimension2, 0]
y = np.array(label)
sample_x = np.append(sample_x, np.array([x]), axis=0)
sample_y = np.append(sample_y, np.array(label))
if label == 0:
counter_sample0 += 1
else:
counter_sample1 += 1
else:
problems = analyze_rpoints(y_signal, rpeaks, sampling_rate)
for p in problems:
problem = int(p)
for index in range(problem, problem + dimension1, step):
if index + dimension1 < rpeaks.shape[0]:
if rpeaks[index + dimension1] + dimension2 < length:
x = np.empty([dimension1, dimension2])
for i in range(dimension1):
x[i] = y_signal[rpeaks[i +
index]:rpeaks[i + index] +
dimension2, 0]
sample_x = np.append(sample_x, np.array([x]), axis=0)
sample_y = np.append(sample_y, np.array(label))
counter_sample1 += 1
sample_y = np.reshape(sample_y, [sample_y.shape[0], 1])
return sample_x, sample_y, counter_label0, counter_label1, counter_sample0, counter_sample1