def predictWithData(self, data, y, _featureDesp="all"):
labels = y
y = [label['value'] for label in labels]
assert (len(data) == len(y))
# split data
num_groundtruth = len(data)
train_size = int(num_groundtruth * self.split_ratio)
test_size = num_groundtruth - train_size
# track test instances in original data for access to metadata
test_glucoseData = self.glucose_data[train_size:]
# assert (len(test_glucoseData) == test_size)
# fix train_size, as we ignored the first value
train_size -= 1
train_data = data[0:train_size]
test_glucoseData = labels[train_size:]
train_y = y[0:train_size]
test_data = data[train_size:]
test_y = y[train_size:]
lr = linear_model.LinearRegression()
lr.fit(train_data, train_y)
predictions = lr.predict(test_data)
if len(test_data) == 0:
return
# self.confidence_cal(train_data, test_data, test_y, predictions, rf, self.patient_id)
results = dict()
results['groundtruth'] = [item['value'] for item in test_glucoseData]
timestamps = [item['time'] for item in test_glucoseData]
results['times'] = timestamps
results['indices'] = [int(item['index']) for item in test_glucoseData]
results['predictions'] = predictions
results['performance'] = compute_performance_time_binned(
timestamps=timestamps, groundtruth=test_y, predictions=predictions)
results['performance'].update(
compute_performance_meals(timestamps=timestamps,
groundtruth=test_y,
predictions=predictions,
carbdata=self.carb_data))
results['params'] = self.save_params()
results['featureDesp'] = _featureDesp
self.plot_learned_model(results['predictions'], results['groundtruth'],
results['times'])
return results
def predict(self):
"""
Runs ContextAVG value prediction.
:return:
"""
assert (self.glucose_data)
# split the data
num_groundtruth = len(self.glucose_data)
train_size = int(num_groundtruth * self.split_ratio)
test_size = num_groundtruth - train_size
test_data = self.glucose_data[train_size:]
assert (len(test_data) == test_size)
# create prediction list using time weighted avg of previous values
predictions = list()
for i in range(0, test_size):
# time of prediction
next_time = test_data[i]['time']
# observed data
prev_glucose = [
item for item in self.glucose_data[:train_size - 1 + i]
]
predictions.append(
self.get_time_weighted_average(prev_glucose, next_time))
assert (len(predictions) == test_size)
# return ground truth (test set) and predictions (as a dict)
test_values = [item['value'] for item in test_data]
results = dict()
results['groundtruth'] = test_values
timestamps = [item['time'] for item in test_data]
results['times'] = timestamps
results['indices'] = [item['index'] for item in test_data]
results['predictions'] = predictions
results['performance'] = compute_performance_time_binned(
timestamps=timestamps,
groundtruth=test_values,
predictions=predictions)
results['performance'].update(
compute_performance_meals(timestamps=timestamps,
groundtruth=test_values,
predictions=predictions,
carbdata=self.carb_data))
results['params'] = None
return results
def predict(self):
"""
Runs last value prediction.
:return:
"""
assert(self.glucose_data)
# split the data
num_groundtruth = len(self.glucose_data)
train_size = int(num_groundtruth * self.split_ratio)
test_size = num_groundtruth - train_size
train_data = self.glucose_data[0:train_size]
test_data = self.glucose_data[train_size:]
assert(len(test_data) == test_size)
# compute avg on training data
last_value = train_data[-1]['value']
# create prediction list using avg
test_values = [item['value'] for item in test_data]
predictions = list()
for i in range(0, test_size):
predictions.append(last_value)
last_value = test_values[i]
assert(len(predictions) == test_size)
# return ground truth (test set) and predictions (as a dict)
results = dict()
results['groundtruth'] = test_values
timestamps = [item['time'] for item in test_data]
results['times'] = timestamps
results['indices'] = [item['index'] for item in test_data]
results['predictions'] = predictions
results['performance'] = compute_performance_time_binned(
timestamps=timestamps,
groundtruth=test_values,
predictions=predictions)
results['performance'].update(compute_performance_meals(
timestamps=timestamps,
groundtruth=test_values,
predictions=predictions,
carbdata=self.carb_data
))
results['params'] = None
return results
def predict(self):
size = int(len(self.discretized_data) * self.split_ratio)
train, test = self.discretized_data[0:size], self.discretized_data[
size:len(self.discretized_data)]
print("number of test instances including nan: {}".format(len(test)))
# remove missing values
train = train[numpy.logical_not(numpy.isnan(train))]
test = test[numpy.logical_not(numpy.isnan(test))]
print("number of test instances: {}".format(len(test)))
history = [x for x in train]
predictions = list()
for t in range(len(test)):
model = ARIMA_MODEL(history, order=(self.p, self.d, self.q))
model_fit = model.fit(disp=0)
output = model_fit.forecast()
yhat = output[0]
predictions.append(yhat)
obs = test[t]
size = int(len(self.glucose_data) * self.split_ratio)
history.append(obs)
print('predicted=%f, expected=%f' % (yhat, obs))
# load timestamps
ts_df = self.load_timestamps(self.con, self.patient_id)
test_ts = ts_df.tail(len(test))
results = dict()
results['groundtruth'] = test
timestamps = test_ts.index
results['times'] = timestamps
results['indices'] = test_ts['pos'].values
results['predictions'] = predictions
results['performance'] = compute_performance_time_binned(
timestamps=timestamps, groundtruth=test, predictions=predictions)
results['performance'].update(
compute_performance_meals(timestamps=timestamps,
groundtruth=test,
predictions=predictions,
carbdata=self.carb_data))
results['params'] = self.save_params()
return results
def to_results(self, test_glucoseData, carbData, test_y, predictions,
classes, patientId):
timestamps = [item['time'] for item in test_glucoseData]
results = dict()
results['groundtruth'] = [item['value'] for item in test_glucoseData]
results['times'] = timestamps
results['indices'] = [int(item['index']) for item in test_glucoseData]
results['performance'], results[
'perClass'] = compute_performance_time_binned(
test_y,
predictions,
timestamps=timestamps,
regression=False,
plotConfusionMatrix=False,
classes=classes,
patientId=patientId,
model=self.modelName)
r_meal, r_meal_perclass = compute_performance_meals(
test_y,
predictions,
timestamps=timestamps,
plotConfusionMatrix=False,
classes=classes,
patientId=patientId,
carbdata=carbData,
regression=False,
model=self.modelName)
results['performance'].update(r_meal)
results['perClass'].update(r_meal_perclass)
results['params'] = self.save_params()
# Compute confusion matrix
cnf_matrix = confusion_matrix(test_y, predictions)
np.set_printoptions(precision=2)
results["report"] = str(cnf_matrix)
return results
def predict(self):
"""
load in data and model_name, whether LSTM or RNN
:return: gt-values and predictions
"""
X, Y = self.load_time_series(
self.con, self.patient_id
) if self.discretized else self.load_continuous_data()
print "modelname %s" % self.modelName
# TODO: fix split when not interpolating
train_size = int(len(X) * self.split_ratio)
test_size = len(X) - train_size
trainX, testX = X[0:train_size], X[train_size:len(X)]
trainY, testY = Y[0:train_size], Y[train_size:len(Y)]
train_size = len(trainY)
test_size = len(testY)
# reshape into X=t and Y=t+1
# TODO: remove the missing values here; unless we want to interpolate: then do it before transforming the data
# scale all the data \o/
xScaler = MinMaxScaler(feature_range=(0, 1))
yScaler = MinMaxScaler(feature_range=(0, 1))
# reshape to 2D for MinMaxScaler
train_nsamples, train_nx, train_ny = trainX.shape
d2_trainX = trainX.reshape((train_nsamples, train_nx * train_ny))
test_nsamples, test_nx, test_ny = testX.shape
d2_testX = testX.reshape((test_nsamples, test_nx * test_ny))
trainX = xScaler.fit_transform(d2_trainX)
testX = xScaler.transform(d2_testX)
trainY = trainY.reshape(-1, 1)
testY = testY.reshape(-1, 1)
trainY = yScaler.fit_transform(trainY)
testY = yScaler.fit_transform(testY)
# reshape input to be [samples, time steps, features]
trainX = numpy.reshape(trainX, (train_nsamples, train_nx, train_ny))
testX = numpy.reshape(testX, (test_nsamples, test_nx, test_ny))
#actual values
testYa = self.get_test_values(testY, train_size, test_size)
# create and fit the LSTM network
model = Sequential()
#FIXED: we skip here the time period features
if self.addTimeofDay: self.no_features += 1
# input shape: (look_back, nfeatures)
#model.add(Masking(mask_value=0., input_shape=(self.look_back, self.no_features)))
#model.add(Model.LSTM.value(4))
model.add(self.models[self.modelName](self.num_units,
input_shape=(self.look_back,
self.no_features)))
model.add(Dense(1))
model.compile(loss='mean_squared_error', optimizer='adam')
model.fit(trainX, trainY, epochs=self.epochs, batch_size=1, verbose=2)
# make predictions
testPredict = model.predict(testX)
print(
"#Test data points (w/ padding): {}; #instances: {}; # training instances: {};"
.format(len(testY), len(X), len(trainX)))
# trace back predictions at actual values
testYp = self.get_test_values(testPredict, train_size, test_size)
testYp = yScaler.inverse_transform(testYp)
testYa = yScaler.inverse_transform(testYa)
print("Test values: {}".format(testYa))
print("predicted values: {}".format(testYp))
# load timestamps
ts_df = self.load_timestamps(self.con, self.patient_id)
# if self.interpolated:
# # skip last timestamp
# ts_df = ts_df[:-1]
test_ts = ts_df.tail(len(testYa))
# test_ts = test_ts.reset_index()
results = dict()
results['groundtruth'] = testYa
timestamps = test_ts.index
results['times'] = timestamps
results['indices'] = test_ts['pos'].values
results['predictions'] = testYp
results['performance'] = compute_performance_time_binned(
timestamps=timestamps, groundtruth=testYa, predictions=testYp)
results['performance'].update(
compute_performance_meals(timestamps=timestamps,
groundtruth=testYa,
predictions=testYp,
carbdata=self.carb_data))
results['params'] = self.save_params()
assert (len(testYa) == len(testYp))
assert (len(timestamps) == len(testYp))
return results
def predict_with_data(self, data, Y, _feature_desp="all"):
# labeling
sorted_Y = sorted(Y) # sort ascending
thresh1 = int(len(sorted_Y) * self.threshold1) - 1
thresh2 = int(len(sorted_Y) * self.threshold2) - 1
if self.multiclass:
cat_Y = [self.categorized_y(y, [sorted_Y[thresh1], sorted_Y[thresh2]]) for y in Y]
classes = ['low ' + str(self.threshold1 * 100) + '%', 'medium', 'high']
else:
cat_Y = [self.categorized_y(y, [sorted_Y[thresh1]]) for y in Y]
classes = ['low ' + str(self.threshold1 * 100) + '%', 'high']
assert (len(data) == len(cat_Y))
# split data
num_groundtruth = len(self.glucose_data)
train_size = int(num_groundtruth * self.split_ratio)
test_size = num_groundtruth - train_size
# track test instances in original data for access to metadata
test_glucose_data = self.glucose_data[train_size:]
assert (len(test_glucose_data) == test_size)
# fix train_size, as we ignored the first value
train_size -= 1
train_data = data[0:train_size]
train_y = cat_Y[0:train_size]
test_data = data[train_size:]
test_y = cat_Y[train_size:]
assert (len(test_y) == len(test_glucose_data))
assert (len(train_y) + len(test_y) + 1 == num_groundtruth)
clf = None
predictions = None
if self.tune:
clf = svm.SVC()
grid = GridSearchCV(clf, param_grid=self.param_grid, cv=5, refit=True, scoring='accuracy')
grid.fit(train_data, train_y)
self.best_params = grid.best_estimator_;
self.log.info("Best parameters for patient {} {}".format(self.patient_id, self.best_params))
predictions = grid.predict(test_data)
else:
clf = svm.SVC()
clf.fit(train_data, train_y)
print train_y
print predictions
predictions = clf.predict(test_data)
print precision_recall_fscore_support(test_y, predictions, average='weighted')
print "accuracy: {}".format(accuracy_score(test_y, predictions, normalize=False))
timestamps = [item['time'] for item in test_glucose_data]
results = dict()
results['performance'], results['perClass'] = compute_performance_time_binned(test_y, predictions,
timestamps=timestamps,
regression=False,
plotConfusionMatrix=True,
classes=classes,
patientId=self.patient_id,
model=self.modelName)
r_meal, r_meal_perclass = compute_performance_meals(test_y, predictions, timestamps=timestamps,
plotConfusionMatrix=True,
classes=classes, patientId=self.patient_id,
carbdata=self.carb_data, regression=False,
model=self.modelName)
results['performance'].update(r_meal)
results['perClass'].update(r_meal_perclass)
results['params'] = self.save_params()
# Compute confusion matrix
cnf_matrix = confusion_matrix(test_y, predictions)
np.set_printoptions(precision=2)
results["report"] = "Binary classification with label: {}".format(classes)
results["report"] = "number of instance in {}: {} and in {}: {}".format(classes[0], thresh1 + 1,
classes[1], len(Y) - thresh1 - 1)
results["report"] += ";confusion matrix: " + str(cnf_matrix)
# Plot non-normalized confusion matrix
save_confusion_matrix(cnf_matrix, classes=classes, patientId=self.patient_id, desc="all", model=self.modelName)
return results
def predict_with_data(self, data, Y, _feature_desp="all"):
# labeling
sorted_Y = sorted(Y) # sort ascending
thresh1 = int(len(sorted_Y) * self.threshold1) - 1
thresh2 = int(len(sorted_Y) * self.threshold2) - 1
'''
TODO: refactor for code reuse at class siblings
'''
if self.hard_threshold:
cat_Y = [
self.categorized_y(y, [Constant.HYPERGLYCEMIA_THRESHOLD])
for y in Y
]
classes = ["non", Constant.HYPER]
else:
if self.multiclass:
cat_Y = [
self.categorized_y(y,
[sorted_Y[thresh1], sorted_Y[thresh2]])
for y in Y
]
classes = [
'low ' + str(self.threshold1 * 100) + '%', 'medium', 'high'
]
else:
cat_Y = [self.categorized_y(y, [sorted_Y[thresh1]]) for y in Y]
classes = ['low ' + str(self.threshold1 * 100) + '%', 'high']
assert (len(data) == len(cat_Y))
# split data
num_groundtruth = len(self.glucose_data)
train_size = int(num_groundtruth * self.split_ratio)
test_size = num_groundtruth - train_size
# track test instances in original data for access to metadata
test_glucose_data = self.glucose_data[train_size:]
assert (len(test_glucose_data) == test_size)
# fix train_size, as we ignored the first value
train_size -= 1
train_data = data[0:train_size]
train_y = cat_Y[0:train_size]
test_data = data[train_size:]
test_y = cat_Y[train_size:]
assert (len(test_y) == len(test_glucose_data))
assert (len(train_y) + len(test_y) + 1 == num_groundtruth)
clf = DummyClassifier(strategy=self.strategy,
random_state=self.random_state)
clf.fit(train_data, train_y) # ignore train_data
predictions = clf.predict(test_data)
print "prediction: {}".format(predictions)
print "accuracy: {}".format(
accuracy_score(test_y, predictions, normalize=True))
print precision_recall_fscore_support(test_y, predictions)
timestamps = [item['time'] for item in test_glucose_data]
results = dict()
results['performance'], results[
'perClass'] = compute_performance_time_binned(
test_y,
predictions,
timestamps=timestamps,
regression=False,
plotConfusionMatrix=True,
classes=classes,
patientId=self.patient_id,
model=self.model_name)
r_meal, r_meal_perclass = compute_performance_meals(
test_y,
predictions,
timestamps=timestamps,
plotConfusionMatrix=True,
classes=classes,
patientId=self.patient_id,
carbdata=self.carb_data,
regression=False,
model=self.model_name)
results['performance'].update(r_meal)
results['perClass'].update(r_meal_perclass)
results['params'] = self.save_params()
# Compute confusion matrix
cnf_matrix = confusion_matrix(test_y, predictions)
np.set_printoptions(precision=2)
results["report"] = "Binary classification with label: {}".format(
classes)
results[
"report"] = "number of instance in {}: {} and in {}: {}".format(
classes[0], thresh1 + 1, classes[1],
len(Y) - thresh1 - 1)
results["report"] += ";confusion matrix: " + str(cnf_matrix)
# Plot non-normalized confusion matrix
save_confusion_matrix(cnf_matrix,
classes=classes,
patientId=self.patient_id,
desc="all",
model=self.model_name)
return results
def predictWithData(self, data, y, _featureDesp="all"):
labels = y
y = [label['value'] for label in labels]
assert (len(data) == len(y))
# split data
num_groundtruth = len(data)
train_size = int(num_groundtruth * self.split_ratio)
test_size = num_groundtruth - train_size
# track test instances in original data for access to metadata
test_glucoseData = self.glucose_data[train_size:]
# assert (len(test_glucoseData) == test_size)
# fix train_size, as we ignored the first value
train_size -= 1
train_data = data[0:train_size]
test_glucoseData = labels[train_size:]
train_y = y[0:train_size]
test_data = data[train_size:]
test_y = y[train_size:]
# assert (len(test_y) == len(test_glucoseData))
# assert (len(train_y) + len(test_y) + 1 == num_groundtruth)
rf = None
if self.tune:
model = self.models[self.modelName](random_state=30)
param_grid = self.param_grid
rf = GridSearchCV(model, param_grid, n_jobs=1, cv=2)
rf.fit(train_data, train_y)
self.best_params = rf.best_estimator_
self.log.info("Best parameters for patient {} {}".format(
self.patient_id, self.best_params))
else:
rf = self.models[self.modelName](
n_estimators=self.n_estimator,
criterion=self.criterion,
min_samples_leaf=self.min_samples_leaf)
# confident intervals with small data
# train_data, train_y = self.sub_data(train_data, train_y)
rf.fit(train_data, train_y)
predictions = rf.predict(test_data)
V_IJ, V_IJ_unbiased = self.confidenceCal(train_data, test_data,
predictions, test_y, rf)
# start = datetime.now()
# predictions = self.feature_select_on_accuracy(train_data, train_y, test_data, rf)
# print "runtime for feature selection for patient{}. : {}".format(self.patient_id, datetime.now() - start)
confidence_thrsd = 0.4 # choose from the parameter tuning
filtered = []
for i in range(0, len(test_data)):
if V_IJ[i] >= confidence_thrsd:
filtered.append(i)
test_data = np.delete(test_data, filtered)
test_y = np.delete(test_y, filtered)
test_glucoseData = np.delete(test_glucoseData, filtered)
predictions = np.delete(predictions, filtered)
if len(test_data) == 0:
return
# self.confidence_cal(train_data, test_data, test_y, predictions, rf, self.patient_id)
results = dict()
results['groundtruth'] = [item['value'] for item in test_glucoseData]
timestamps = [item['time'] for item in test_glucoseData]
results['times'] = timestamps
results['indices'] = [int(item['index']) for item in test_glucoseData]
results['predictions'] = predictions
results['performance'] = compute_performance_time_binned(
timestamps=timestamps, groundtruth=test_y, predictions=predictions)
results['performance'].update(
compute_performance_meals(timestamps=timestamps,
groundtruth=test_y,
predictions=predictions,
carbdata=self.carb_data))
results['params'] = self.save_params()
results['featureDesp'] = _featureDesp
self.plot_learned_model(results['predictions'], results['groundtruth'],
results['times'])
return results
def predict_with_data(self, data, Y, _feature_desp="all"):
# labeling
sorted_Y = sorted(Y) # sort ascending
thresh1 = int(len(sorted_Y) * self.threshold1) - 1
thresh2 = int(len(sorted_Y) * self.threshold2) - 1
'''
TODO: refactor for code reuse at class siblings
'''
if self.hard_threshold:
cat_Y = [self.categorized_y(y, [Constant.HYPERGLYCEMIA_THRESHOLD]) for y in Y]
classes = ["non", Constant.HYPER]
else:
if self.multiclass:
cat_Y = [self.categorized_y(y, [sorted_Y[thresh1], sorted_Y[thresh2]]) for y in Y]
classes = ['low ' + str(self.threshold1 * 100) + '%', 'medium', 'high']
else:
cat_Y = [self.categorized_y(y, [sorted_Y[thresh1]]) for y in Y]
classes = ['low ' + str(self.threshold1 * 100) + '%', 'high']
assert (len(data) == len(cat_Y))
# split data
num_groundtruth = len(self.glucose_data)
train_size = int(num_groundtruth * self.split_ratio)
test_size = num_groundtruth - train_size
# track test instances in original data for access to metadata
test_glucose_data = self.glucose_data[train_size:]
assert (len(test_glucose_data) == test_size)
# fix train_size, as we ignored the first value
train_size -= 1
train_data = data[0:train_size]
train_y = cat_Y[0:train_size]
test_data = data[train_size:]
test_y = cat_Y[train_size:]
assert (len(test_y) == len(test_glucose_data))
assert (len(train_y) + len(test_y) + 1 == num_groundtruth)
rf = None
if self.tune:
model = ensemble.RandomForestClassifier(random_state=30, class_weight="balanced")
param_grid = self.param_grid
rf = GridSearchCV(model, param_grid, n_jobs=-1, cv=2)
self.best_params = rf.best_estimator_;
self.log.info("Best parameters for patient %s %s") % self.patient_id % self.best_params
else:
rf = self.models[self.modelName](n_estimators=self.n_estimator, criterion=self.criterion,
min_samples_leaf=self.min_samples_leaf, class_weight="balanced")
'''
Sampling training set
'''
print "Original training set shape: {}".format(Counter(train_y))
print "Original test set shape: {}".format(Counter(test_y))
self.log.debug("Original training set shape: {}".format(Counter(train_y)))
# try:
# sm = SMOTE(random_state=42, k_neighbors=3)
# train_data, train_y = sm.fit_sample(np.asarray(train_data), np.asarray(train_y))
# self.log.debug("Resampled training set shape: {}".format(Counter(train_y)))
# except ValueError:
# pass
rf.fit(train_data, train_y)
predictions = rf.predict(test_data)
print "prediction: {}".format(predictions)
print "accuracy: {}".format(accuracy_score(test_y, predictions, normalize=True))
print precision_recall_fscore_support(test_y,predictions)
timestamps = [item['time'] for item in test_glucose_data]
results = dict()
results['groundtruth'] = [item['value'] for item in test_glucose_data]
results['times'] = timestamps
results['indices'] = [int(item['index']) for item in test_glucose_data]
results['performance'], results['perClass'] = compute_performance_time_binned(test_y, predictions, timestamps=timestamps,
regression=False, plotConfusionMatrix=False,
classes=classes, patientId=self.patient_id,
model=self.modelName)
r_meal, r_meal_perclass = compute_performance_meals(test_y, predictions, timestamps=timestamps, plotConfusionMatrix=False,
classes=classes, patientId=self.patient_id, carbdata=self.carb_data, regression=False,
model=self.modelName)
results['performance'].update(r_meal)
results['perClass'].update(r_meal_perclass)
results['params'] = self.save_params()
# Compute confusion matrix
cnf_matrix = confusion_matrix(test_y, predictions)
np.set_printoptions(precision=2)
results["report"] = "Binary classification with label: {}".format(classes)
results["report"] = "number of instance in {}: {} and in {}: {}".format(classes[0], thresh1 + 1,
classes[1], len(Y) - thresh1 - 1)
results["report"] += ";confusion matrix: " + str(cnf_matrix)
#Plot non-normalized confusion matrix
save_confusion_matrix(cnf_matrix, classes=classes, patientId=self.patient_id, desc="all", model=self.modelName)
return results