def get_feature_names():
id_train_all, x_train_all, y_train_all = data_util.get_poor_god(
'training_is_0.csv', sub_class='ischemic')
x_train_all.rename(columns={
'MRS_TX_1': '30-day mRS',
'discharged_mrs': 'Discharge mRS',
'Toilet_use': 'Toilet use',
'Bowel_control': 'Bowel control',
'Bladder_control': 'Bladder control',
'TRMNG_FL': 'Nasogastric tube',
'TRMRE_FL': 'Rehab',
'OFFDT_ID_1': 'Discharge to Home',
'NIHS_6aL_out': 'Discharge NIHSS 6aL',
'NIHS_6aL_in': 'Admission NIHSS 6aL',
'NIHS_6bR_out': 'Discharge NIHSS 6bR',
'NIHS_10_out': 'Discharge NIHSS 10',
'NIHS_5aL_out': 'Discharge NIHSS 5aL',
'NIHS_5bR_out': 'Discharge NIHSS 5bR',
'NIHS_1b_out': 'Discharge NIHSS 1b',
'NIHS_9_out': 'Discharge NIHSS 9',
'NIHS_5aL_in': 'Admission NIHSS 5aL',
'NIHS_1b_in': 'Admission NIHSS 1b'
},
inplace=True)
return x_train_all.columns.values
test_loss_array = []
predict_array = []
# ====== Multi-classes
# x_data, y_data = data_util.get_individual('wholeset_Jim_nomissing_validated.csv')
# for index, (train, test) in enumerate(kfold.split(x_data, y_data)):
# history, model = mlp_multi(data_util.scale(x_data.iloc[train]),
# to_categorical(y_data.iloc[train]),
# parameter)
# history_array.append(history)
# loss, acc = model.evaluate(data_util.scale(x_data.iloc[test]),
# to_categorical(y_data.iloc[test]),
# verbose=0)
# test_acc_array.append(acc)
# test_loss_array.append(loss)
# ====== Binary
id_data, x_data, y_data = data_util.get_poor_god(
'wholeset_Jim_nomissing_validated.csv')
for index, (train, test) in enumerate(kfold.split(x_data, y_data)):
x_train_cnn, x_train_mlp = data_util.split_cnn_mlp_input(
x_data.iloc[train])
# x_train = data_util.kera_feature(x_data.iloc[train])
history, model = mlp_binary(data_util.scale(x_train_cnn),
to_categorical(y_data.iloc[train]),
parameter)
history_array.append(history)
x_test_cnn, x_test_mlp = data_util.split_cnn_mlp_input(
x_data.iloc[test])
# x_test = data_util.kera_feature(x_data.iloc[test])
loss, acc = model.evaluate(data_util.scale(x_test_cnn),
to_categorical(y_data.iloc[test]),
verbose=0)
import numpy as np
import pandas as pd
import os
from sklearn.ensemble import ExtraTreesClassifier
from sklearn.model_selection import StratifiedKFold
from my_utils import data_util
if __name__ == '__main__':
subtype = 'he'
# hold_out_round = 1
for hold_out_round in range(0, 10, 1):
if subtype == 'is':
sub_class = 'ischemic'
else:
sub_class = 'hemorrhagic'
id_train_all, x_train_all, y_train_all = data_util.get_poor_god('training_' + subtype + '_' + str(hold_out_round) + '.csv', sub_class=sub_class)
feature_names = x_train_all.columns.values
forest = ExtraTreesClassifier()
kfold = StratifiedKFold(n_splits=10, shuffle=True, random_state=hold_out_round)
for index, (train, test) in enumerate(kfold.split(x_train_all, y_train_all)):
x_train = data_util.scale(x_train_all.iloc[train])
y_train = y_train_all.iloc[train]
forest.fit(x_train, y_train)
importances = forest.feature_importances_
std = np.std([tree.feature_importances_ for tree in forest.estimators_], axis=0)
indices = np.argsort(importances)[::-1]
# Print the feature ranking
imp = []
print("Feature ranking:")
for i in range(x_train.shape[1]):
print("%d. feature %d (%f) %s" % (i + 1, indices[i], importances[indices[i]], feature_names[indices[i]]))
import pandas as pd
import os
import numpy as np
import matplotlib.pyplot as plt #Data visualisation libraries
import seaborn as sns
from sklearn.linear_model import LogisticRegression
from my_utils import data_util, performance_util
from sklearn.metrics import roc_auc_score
id_train_all, x_train_all, y_train_all = data_util.get_poor_god_downsample(
'training_is_9.csv', sub_class='ischemic')
id_hold, x_hold, y_hold = data_util.get_poor_god('hold_is_9.csv',
sub_class='ischemic')
lm = LogisticRegression()
x = x_train_all[['MRS_TX_1']]
y = y_train_all
lm.fit(x, y)
test_x = x_hold[['MRS_TX_1']]
predictions = lm.predict(test_x)
logit_roc_auc = roc_auc_score(y_hold, lm.predict(test_x))
print(logit_roc_auc)
def do_svm(hold_out_round, sub_class, experiment):
np.random.seed(hold_out_round)
if sub_class == 'ischemic':
id_train_all, x_train_all, y_train_all = data_util.get_poor_god_downsample(
'training_is_' + str(hold_out_round) + '.csv', sub_class=sub_class)
id_hold, x_hold, y_hold = data_util.get_poor_god(
'hold_is_' + str(hold_out_round) + '.csv', sub_class=sub_class)
else:
id_train_all, x_train_all, y_train_all = data_util.get_poor_god_downsample(
'training_he_' + str(hold_out_round) + '.csv', sub_class=sub_class)
id_hold, x_hold, y_hold = data_util.get_poor_god(
'hold_he_' + str(hold_out_round) + '.csv', sub_class=sub_class)
#
if experiment == 0:
save_path = '..' + os.sep + 'result' + os.sep + 'svm' + os.sep + 'all' + os.sep
model_name = 'svm_' + sub_class + '_h_' + str(hold_out_round)
elif experiment == 1:
x_train_all = data_util.feature_selection(x_train_all, sub_class)
x_hold = data_util.feature_selection(x_hold, sub_class)
save_path = '..' + os.sep + 'result' + os.sep + 'svm' + os.sep + 'fs' + os.sep
model_name = 'svm_fs_' + sub_class + '_h_' + str(hold_out_round)
elif experiment == 2:
x_train_all = x_train_all.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_hold = x_hold.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
save_path = '..' + os.sep + 'result' + os.sep + 'svm' + os.sep + 'all_nf' + os.sep
model_name = 'svm_nf_' + sub_class + '_h_' + str(hold_out_round)
else:
x_train_all = data_util.feature_selection(x_train_all, sub_class)
x_train_all = x_train_all.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_hold = data_util.feature_selection(x_hold, sub_class)
x_hold = x_hold.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
save_path = '..' + os.sep + 'result' + os.sep + 'svm' + os.sep + 'fs_nf' + os.sep
model_name = 'svm_fs_nf_' + sub_class + '_h_' + str(hold_out_round)
#
test_acc_array = []
kfold = StratifiedKFold(n_splits=10,
shuffle=True,
random_state=hold_out_round)
classifier = SVC(kernel='linear',
probability=True,
random_state=hold_out_round,
verbose=True)
for index, (train, test) in enumerate(kfold.split(x_train_all,
y_train_all)):
# Training
x_train = data_util.scale(x_train_all.iloc[train])
y_train = y_train_all.iloc[train]
# Testing
x_test = data_util.scale(x_train_all.iloc[test])
y_test = y_train_all.iloc[test]
# train on 90% training
classifier.fit(x_train, y_train)
predict_result_train = id_train_all.iloc[train]
train_probas = classifier.predict_proba(x_train)
predict_result_train['label'] = y_train
predict_result_train['0'] = train_probas[:, 0]
predict_result_train['1'] = train_probas[:, 1]
predict_result_train.to_csv(save_path + model_name + '_train_cv' +
str(index) + '.csv',
sep=',',
encoding='utf-8')
# Evaluation on 10% training
predict_result_test = id_train_all.iloc[test]
test_probas = classifier.predict_proba(x_test)
predict_result_test['label'] = y_test
predict_result_test['0'] = test_probas[:, 0]
predict_result_test['1'] = test_probas[:, 1]
predict_result_test.to_csv(save_path + model_name + '_test_cv' +
str(index) + '.csv',
sep=',',
encoding='utf-8')
test_acc = accuracy_score(y_test, classifier.predict(x_test))
test_acc_array.append(test_acc)
performance_util.save_model(classifier, model_name + '_' + str(index))
print('10-CV Done')
# --
best_model_inx = test_acc_array.index(max(test_acc_array))
hold_model = performance_util.load_ml_model(model_name, best_model_inx)
x_hold = data_util.scale(x_hold)
predict_result_hold = id_hold
holdout_probas = hold_model.predict_proba(x_hold)
predict_result_hold['label'] = y_hold
predict_result_hold['0'] = holdout_probas[:, 0]
predict_result_hold['1'] = holdout_probas[:, 1]
predict_result_hold.to_csv(save_path + model_name + '_hold.csv',
sep=',',
encoding='utf-8')
print('hold-out Done')
def do_mlp_cnn(hold_out_round, sub_class, experiment):
np.random.seed(hold_out_round)
if sub_class == 'ischemic':
id_train_all, x_train_all, y_train_all = data_util.get_poor_god_downsample(
'training_is_' + str(hold_out_round) + '.csv', sub_class=sub_class)
id_hold, x_hold, y_hold = data_util.get_poor_god(
'hold_is_' + str(hold_out_round) + '.csv', sub_class=sub_class)
else:
id_train_all, x_train_all, y_train_all = data_util.get_poor_god_downsample(
'training_he_' + str(hold_out_round) + '.csv', sub_class=sub_class)
id_hold, x_hold, y_hold = data_util.get_poor_god(
'hold_he_' + str(hold_out_round) + '.csv', sub_class=sub_class)
#
if experiment == 0:
save_path = '..' + os.sep + 'result' + os.sep + 'mlp_cnn' + os.sep + 'all' + os.sep
parameter = {
'model_name': 'mlp_cnn_' + sub_class + '_h_' + str(hold_out_round),
'size_of_batch': 56,
'nb_epoch': 150,
'drop_rate': 0.5
}
elif experiment == 1:
save_path = '..' + os.sep + 'result' + os.sep + 'mlp_cnn' + os.sep + 'fs' + os.sep
selected_features = data_util.get_selected_feature_name(sub_class)
parameter = {
'model_name':
'mlp_cnn_fs_' + sub_class + '_h_' + str(hold_out_round),
'size_of_batch': 56,
'nb_epoch': 150,
'drop_rate': 0.5
}
elif experiment == 2:
save_path = '..' + os.sep + 'result' + os.sep + 'mlp_cnn' + os.sep + 'all_nf' + os.sep
parameter = {
'model_name':
'mlp_cnn_nf_' + sub_class + '_h_' + str(hold_out_round),
'size_of_batch': 56,
'nb_epoch': 150,
'drop_rate': 0.5
}
else:
save_path = '..' + os.sep + 'result' + os.sep + 'mlp_cnn' + os.sep + 'fs_nf' + os.sep
selected_features = data_util.get_selected_feature_name(sub_class)
parameter = {
'model_name':
'mlp_cnn_fs_nf_' + sub_class + '_h_' + str(hold_out_round),
'size_of_batch':
56,
'nb_epoch':
150,
'drop_rate':
0.5
}
test_acc_array = []
test_loss_array = []
kfold = StratifiedKFold(n_splits=10,
shuffle=True,
random_state=hold_out_round)
for index, (train, test) in enumerate(kfold.split(x_train_all,
y_train_all)):
# training
x_train_cnn, x_train_mlp = data_util.split_cnn_mlp_input(
x_train_all.iloc[train])
if experiment == 2:
x_train_cnn = x_train_cnn.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_train_mlp = x_train_mlp.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
if experiment == 1 or experiment == 3:
x_train_cnn, x_train_mlp = data_util.selected_cnn_mlp_input(
x_train_cnn, x_train_mlp, selected_features)
if experiment == 3:
x_train_cnn = x_train_cnn.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0',
'FLU_ID_1_4.0', 'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1',
'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_train_mlp = x_train_mlp.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0',
'FLU_ID_1_4.0', 'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1',
'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_train_cnn = data_util.scale(x_train_cnn)
x_train_mlp = data_util.scale(x_train_mlp)
y_train = y_train_all.iloc[train]
# Testing
x_test_cnn, x_test_mlp = data_util.split_cnn_mlp_input(
x_train_all.iloc[test])
if experiment == 2:
x_test_cnn = x_test_cnn.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_test_mlp = x_test_mlp.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
if experiment == 1 or experiment == 3:
x_test_cnn, x_test_mlp = data_util.selected_cnn_mlp_input(
x_test_cnn, x_test_mlp, selected_features)
if experiment == 3:
x_test_cnn = x_test_cnn.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0',
'FLU_ID_1_4.0', 'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1',
'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_test_mlp = x_test_mlp.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0',
'FLU_ID_1_4.0', 'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1',
'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_test_cnn = np.expand_dims(data_util.scale(x_test_cnn), 2)
x_test_mlp = data_util.scale(x_test_mlp)
y_test = y_train_all.iloc[test]
# train on 90% training
history, model = mlp_cnn_binary(x_train_cnn, x_train_mlp,
to_categorical(y_train), parameter,
index)
performance_util.save_train_validation(
save_path + parameter['model_name'], history, 'acc', str(index))
predict_result_train = id_train_all.iloc[train]
x_train_cnn = np.expand_dims(x_train_cnn, 2)
train_probas = model.predict([x_train_cnn, x_train_mlp])
predict_result_train['label'] = y_train
predict_result_train['0'] = train_probas[:, 0]
predict_result_train['1'] = train_probas[:, 1]
predict_result_train.to_csv(save_path + parameter['model_name'] +
'_train_cv' + str(index) + '.csv',
sep=',',
encoding='utf-8')
# Evaluation on 10% training
predict_result_test = id_train_all.iloc[test]
test_probas = model.predict([x_test_cnn, x_test_mlp])
predict_result_test['label'] = y_test
predict_result_test['0'] = test_probas[:, 0]
predict_result_test['1'] = test_probas[:, 1]
predict_result_test.to_csv(save_path + parameter['model_name'] +
'_test_cv' + str(index) + '.csv',
sep=',',
encoding='utf-8')
loss, acc = model.evaluate([x_test_cnn, x_test_mlp],
to_categorical(y_test),
verbose=0)
test_acc_array.append(acc)
test_loss_array.append(loss)
# plot_fig.plot_acc_loss(history, 'acc')
performance_util.save_test(save_path + parameter['model_name'],
test_acc_array, test_loss_array)
print('10-CV Done')
# --
best_model_inx = test_acc_array.index(max(test_acc_array))
hold_model = performance_util.load_nn_model(parameter['model_name'],
best_model_inx)
x_hold_cnn, x_hold_mlp = data_util.split_cnn_mlp_input(x_hold)
if experiment == 2:
x_hold_cnn = x_hold_cnn.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_hold_mlp = x_hold_mlp.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
if experiment == 1 or experiment == 3:
x_hold_cnn, x_hold_mlp = data_util.selected_cnn_mlp_input(
x_hold_cnn, x_hold_mlp, selected_features)
if experiment == 3:
x_hold_cnn = x_hold_cnn.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_hold_mlp = x_hold_mlp.drop([
'FLU_ID_1_1.0', 'FLU_ID_1_2.0', 'FLU_ID_1_3.0', 'FLU_ID_1_4.0',
'FLU_ID_1_5.0', 'FLU_ID_1_6.0', 'VERS_1', 'VEIHD_1', 'MRS_TX_1'
],
errors='ignore',
axis=1)
x_hold_cnn = np.expand_dims(data_util.scale(x_hold_cnn), 2)
x_hold_mlp = data_util.scale(x_hold_mlp)
predict_result_hold = id_hold
holdout_probas = hold_model.predict([x_hold_cnn, x_hold_mlp])
predict_result_hold['label'] = y_hold
predict_result_hold['0'] = holdout_probas[:, 0]
predict_result_hold['1'] = holdout_probas[:, 1]
predict_result_hold.to_csv(save_path + parameter['model_name'] +
'_hold.csv',
sep=',',
encoding='utf-8')
print('hold-out Done')
else:
all_selected_features = np.append(all_selected_features,
selected_f_names)
feature_dict = Counter(all_selected_features)
# use to draw heatmap
return list(feature_dict.keys())
if __name__ == '__main__':
# Just get the feature names
subtype = 'he'
if subtype == 'is':
sub_class = 'ischemic'
else:
sub_class = 'hemorrhagic'
id_train_all, x_train_all, y_train_all = data_util.get_poor_god(
'training_' + subtype + '_0.csv', sub_class=sub_class)
feature_names = x_train_all.columns.values
#
for hold_out_round in range(0, 10, 1):
if subtype == 'is':
sub_class = 'ischemic'
else:
sub_class = 'hemorrhagic'
df = pd.read_csv('f_' + subtype + '_' + str(hold_out_round) + '.csv',
encoding='utf8')
mean_importance = df.drop(['f_index'], axis=1).mean(axis=1)
if hold_out_round == 0:
robust_f_df = pd.DataFrame(data={'f_index': df['f_index']})
robust_f_df['rf' + str(hold_out_round)] = mean_importance
else:
robust_f_df['rf' + str(hold_out_round)] = mean_importance