def train_coxph(data_df, r_splits):
c_index_at = []
c_index_30 = []
time_auc_30 = []
time_auc_60 = []
time_auc_365 = []
for i in range(len(r_splits)):
print("\nIteration %s"%(i))
#DATA PREP
df_train, df_val, df_test, df_test_30 = prepare_datasets(data_df, r_splits[i][2], r_splits[i][1], r_splits[i][0])
(data_x, data_y), (val_x, val_y), (test_x, test_y), (test_30_x, test_30_y) = df2array(data_df, df_train, df_val, df_test, df_test_30)
estimator = CoxPHSurvivalAnalysis(alpha=1e-04)
estimator.fit(data_x, data_y)
c_index_at.append(estimator.score(test_x, test_y))
c_index_30.append(estimator.score(test_30_x, test_30_y))
for time_x in [30, 60, 365]:
t_auc, t_mean_auc = cumulative_dynamic_auc(data_y, test_y, estimator.predict(test_x), time_x)
eval("time_auc_" + str(time_x)).append(t_auc[0])
print("C-index_30:", c_index_30[i])
print("C-index_AT:", c_index_at[i])
print("time_auc_30", time_auc_30[i])
print("time_auc_60", time_auc_60[i])
print("time_auc_365", time_auc_365[i])
return c_index_at, c_index_30, time_auc_30, time_auc_60, time_auc_365
def mpss_ph_sksurv(self):
"""
Performs proportional hazards regression using sksurv package.
:return: Feature importance
"""
# Reformat for sksurv package
x_train = pd.DataFrame(self.x_train)
y_structured = [
(ll, s) for ll, s in zip(self.y_train.astype(bool), self.scores)
]
y_structured = np.array(y_structured,
dtype=[('class', 'bool_'),
('score', 'single')])
# Remove any feature columns that are all 0 values, otherwise cannot run regression
x_train_nonzero = x_train.loc[:, (x_train != 0).any(axis=0)]
# Run proportional hazards regression
estimator = CoxPHSurvivalAnalysis(alpha=0.1, verbose=1)
estimator.fit(x_train_nonzero, y_structured)
prediction = estimator.predict(x_train_nonzero)
# Estimate p-values for each feature
f, pvals = f_regression(x_train_nonzero, [x[1] for x in y_structured])
approximate_se = pd.DataFrame(pd.Series(
pvals, index=x_train_nonzero.columns).sort_values(ascending=False),
columns=['p']).reset_index()
# Calculate concordance indicating the goodness of fit
concordance = concordance_index_censored(self.y_train.astype(bool),
self.scores, prediction)
print('concordance', concordance[0])
# Dataframe with coefficients, absolute value of coefficients, and p-values
importance = pd.DataFrame(estimator.coef_, columns=['coef'])
importance['coef_abs'] = [math.fabs(c) for c in importance['coef']]
importance['feature'] = importance.index.values
importance = importance.merge(approximate_se,
left_on='feature',
right_on='index').drop('index', axis=1)
# Sort feature importance
importance = importance.sort_values(
'coef_abs', ascending=False).reset_index(drop=True)
return importance
def cox(name):
filename = filename_dict[name]
raw_data = pd.read_csv(os.path.join(DATA_DIR, filename))
formatted_x, formatted_y = sksurv_data_formatting(raw_data)
x_train, x_test, y_train, y_test = train_test_split(
formatted_x, formatted_y, test_size=0.25, random_state=RANDOM_STATE)
estimator = CoxPHSurvivalAnalysis()
estimator.fit(x_train, y_train)
prediction = estimator.predict(x_test)
result = concordance_index_censored(y_test["Status"],
y_test["Survival_in_days"], prediction)
return result[0]
import numpy as np
pred_surv = estimator.predict_survival_function(x_new)
time_points = np.arange(1, 1000)
for i, surv_func in enumerate(pred_surv):
plt.step(time_points, surv_func(time_points), where="post",
label="Sample %d" % (i + 1))
plt.ylabel("est. probability of survival $\hat{S}(t)$")
plt.xlabel("time $t$")
plt.legend(loc="best")
from sksurv.metrics import concordance_index_censored
prediction = estimator.predict(data_x_numeric)
result = concordance_index_censored(data_y["Status"], data_y["Survival_in_days"], prediction)
result[0]
estimator.score(data_x_numeric, data_y)
# Feature selection
import numpy as np
def fit_and_score_features(X, y):
n_features = X.shape[1]
scores = np.empty(n_features)
m = CoxPHSurvivalAnalysis()
for j in range(n_features):
Xj = X[:, j:j+1]
for f_name in selected_features[n_split]
])
x = np.swapaxes(x, 0, 1) # (n_samples, n_features)
x_train, x_test = x[split['train']], x[split['test']]
y_train, y_test = survial_data[
split['train']], survial_data[split['test']]
# Model
predictor = CoxPHSurvivalAnalysis(alpha=0, n_iter=1e9)
try:
predictor.fit(x_train, y_train[['event', 'time']])
c_indexes.append(
predictor.score(x_test, y_test[['event', 'time']]))
risk_score_train = predictor.predict(x_train)
risk_score = predictor.predict(x_test)
high_risk_masks.append(
risk_score > np.median(risk_score_train))
y_tests.append(y_test)
except Exception as e:
logger.warning("Error {}".format(str(e)))
c_indexes.append(np.NaN)
# ----------------------- Kaplan-Meier --------------------------------
high_risk_mask = np.concatenate(high_risk_masks)
y_tests = np.concatenate(y_tests)
y_high_risk, y_low_risk = y_tests[high_risk_mask], y_tests[
~high_risk_mask]
'regressor__model__min_child_weight': (10, 500, 'log-uniform'), # categorical parameter
'regressor__model__n_estimators': (1, 8), # integer valued parameter
'regressor__model__reg_alpha': (1, 8, 'log-uniform'), # integer valued parameter
'regressor__model__reg_lambda': (1, 8, 'log-uniform'), # integer valued parameter
'regressor__model__subsample': (1, 8, 'log-uniform'), # integer valued parameter
}
#%%
# Since sksurv output log hazard ratios (here relative to 0 on predictors)
# we must use 'output_margin=True' for comparability.
estimator = CoxPHSurvivalAnalysis().fit(data_x, data_y)
gbm = xgb.XGBRegressor(objective='survival:cox',
booster='gblinear',
base_score=1,
n_estimators=1000)
search = BayesSearchCV(gbm, params, n_iter=3, cv=3)
search.fit(data_x, data_y_xgb)
#%%
prediction_sksurv = estimator.predict(data_x)
predictions_xgb = search.predict(data_x)
d = pd.DataFrame({'xgb': predictions_xgb,
'sksurv': prediction_sksurv})
d.head()
# %%
context.io.save('xente_xgb', gbm)
# %%
