from util import compute_median_survival_time, compute_mean_survival_time
def survival_time_nonconformity(predicted_survival_time,
observed_time, event_ind):
if event_ind == 1:
return np.abs(observed_time - predicted_survival_time)
else:
return np.maximum(observed_time - predicted_survival_time, 0)
mean_predicted_survival_times = \
np.array([compute_mean_survival_time(time_points, row) for row in surv])
median_predicted_survival_times = \
np.array([compute_median_survival_time(time_points, row) for row in surv])
for survival_time_estimator in ['mean', 'median']:
if survival_time_estimator == 'mean':
predicted_survival_times = mean_predicted_survival_times.copy()
else:
predicted_survival_times = median_predicted_survival_times.copy()
for calib_frac in np.linspace(0.1, 1, 10):
print('[Split conformal prediction - %s, calib frac %f]'
% (survival_time_estimator, calib_frac))
rng = np.random.RandomState(conformal_prediction_random_seed)
coverages = {conformal_prediction_CI_coverage: []
kernel_function=kernel_func)
cum_hazard_scores = cum_haz.sum(axis=1)
cindex = concordance_index(fold_y_val[:, 0],
-cum_hazard_scores,
fold_y_val[:, 1])
elif cindex_method == 'cum_haz_from_surv':
surv_thresholded = np.maximum(surv,
np.finfo(float).eps)
cum_haz = -np.log(surv_thresholded)
cum_hazard_scores = cum_haz.sum(axis=1)
cindex = concordance_index(fold_y_val[:, 0],
-cum_hazard_scores,
fold_y_val[:, 1])
elif cindex_method == 'median':
predicted_medians = \
np.array([compute_median_survival_time(mesh_points,
surv_row)
for surv_row in surv])
cindex = concordance_index(fold_y_val[:, 0],
predicted_medians,
fold_y_val[:, 1])
elif cindex_method == 'median_from_cum_haz':
cum_haz = \
surv_model.predict_cum_haz(
fold_X_val_standardized,
sorted_fold_y_val,
presorted_times=True,
kernel_function=kernel_func)
predicted_medians = \
np.array([compute_median_survival_time(mesh_points,
surv_row)
for surv_row in np.exp(-cum_haz)])
