def predict_cumulative_hazard(self, x, t=None, **kwargs):
""" Predicts the cumulative hazard function H(t, x)
Parameters
----------
* `x` : **array-like** *shape=(n_samples, n_features)* --
array-like representing the datapoints.
x should not be standardized before, the model
will take care of it
* `t`: **double** *(default=None)* --
time at which the prediction should be performed.
If None, then return the function for all available t.
Returns
-------
* `cumulative_hazard`: **numpy.ndarray** --
array-like representing the prediction of the cumulative_hazard
function
"""
# Checking if the data has the right format
x = utils.check_data(x)
# Calculating hazard/cumulative_hazard
hazard = self.predict_hazard(x, t, **kwargs)
cumulative_hazard = np.cumsum(hazard, 1)
return cumulative_hazard
def predict(self, X, t=None, num_threads=-1):
# Checking if the data has the right format
X = utils.check_data(X)
if X.ndim == 1:
X = X.reshape(1, -1)
T = np.array([1.] * X.shape[0])
E = np.array([1.] * X.shape[0])
input_data = np.c_[T, E, X]
# Loading the attributes of the model
self.load_properties()
# Computing Survival
survival = np.array(
self.model.predict_survival(input_data, num_threads))
# Computing hazard
hazard = np.array(self.model.predict_hazard(input_data, num_threads))
# Computing density
density = hazard * survival
if t is None:
return hazard, density, survival
else:
min_index = [abs(a_j_1 - t) for (a_j_1, a_j) in self.time_buckets]
index = np.argmin(min_index)
return hazard[:, index], density[:, index], survival[:, index]
def predict_risk(self, x, **kwargs):
""" Predicts the Risk Score/Mortality function for all t,
R(x) = sum( cumsum(hazard(t, x)) )
According to Random survival forests from Ishwaran H et al
https://arxiv.org/pdf/0811.1645.pdf
Parameters
----------
* `x` : **array-like** *shape=(n_samples, n_features)* --
array-like representing the datapoints.
x should not be standardized before, the model
will take care of it
Returns
-------
* `risk_score`: **numpy.ndarray** --
array-like representing the prediction of Risk Score function
"""
# Checking if the data has the right format
x = utils.check_data(x)
# Calculating cumulative_hazard/risk
cumulative_hazard = self.predict_cumulative_hazard(x, None, **kwargs)
risk_score = np.sum(cumulative_hazard, 1)
return risk_score
def predict_risk_chunk(self, X, num_threads=-1, printChunk=False):
# Checking if the data has the right format
X = utils.check_data(X)
if X.ndim == 1:
X = X.reshape(1, -1)
T = np.array([1.] * X.shape[0])
E = np.array([1.] * X.shape[0])
input_data = np.c_[T, E, X]
# Loading the attributes of the model
self.load_properties()
n_rows, n_cols = input_data.shape
nparts = round(n_rows / 1000)
split_array = np.array_split(input_data, nparts)
survival = np.empty(n_rows)
survIndex = 0
for s in split_array:
survival[survIndex:(survIndex + np.size(s, axis=0))] = np.array(
self.model.predict_risk(s, num_threads))
survIndex += np.size(s, axis=0)
#if printChunk:
#print(survIndex)
# Computing risk
return survival
def predict(self, x, t=None):
"""
Predicting the hazard, density and survival functions
Arguments:
* x: pd.Dataframe or np.ndarray or list
x is the testing dataset containing the features
x should not be standardized before, the model
will take care of it
* t: float (default=None)
Time at which hazard, density and survival functions
should be calculated. If None, the method returns
the functions for all times t.
"""
# Convert x into the right format
x = utils.check_data(x)
# Sacling the dataset
if x.ndim == 1:
x = self.scaler.transform(x.reshape(1, -1))
elif x.ndim == 2:
x = self.scaler.transform(x)
# Calculating risk_score, hazard, density and survival
phi = np.exp(np.dot(x, self.weights))
hazard = self.baseline_hazard * phi.reshape(-1, 1)
survival = np.power(self.baseline_survival, phi.reshape(-1, 1))
density = hazard * survival
if t is None:
return hazard, density, survival
else:
min_index = [abs(a_j_1 - t) for (a_j_1, a_j) in self.time_buckets]
index = np.argmin(min_index)
return hazard[:, index], density[:, index], survival[:, index]
def predict_risk(self, x):
"""
Predicting the risk score function
Parameters:
-----------
* x: pd.Dataframe or np.ndarray or list
x is the testing dataset containing the features
x should not be standardized before, the model
will take care of it
"""
# Convert x into the right format
x = utils.check_data(x)
# Scaling the dataset
if x.ndim == 1:
x = self.scaler.transform(x.reshape(1, -1))
elif x.ndim == 2:
x = self.scaler.transform(x)
else:
# Ensuring x has 2 dimensions
if x.ndim == 1:
x = np.reshape(x, (1, -1))
# Calculating risk_score
risk_score = self.risk_function(x)
return risk_score
def predict_risk(self, x, use_log=False):
"""
Predicting the risk score functions
Arguments:
* x: pd.Dataframe or np.ndarray or list
x is the testing dataset containing the features
x should not be standardized before, the model
will take care of it
"""
# Convert x into the right format
x = utils.check_data(x)
# Scaling the dataset
if x.ndim == 1:
x = self.scaler.transform(x.reshape(1, -1))
elif x.ndim == 2:
x = self.scaler.transform(x)
# Calculating risk_score
risk_score = np.exp(np.dot(x, self.weights))
if not use_log:
risk_score = np.exp(risk_score)
return risk_score
def predict_risk(self, x, use_log=False):
"""
Predicting the risk score functions
Arguments:
* x: pd.Dataframe or np.ndarray or list
x is the testing dataset containing the features
x should not be standardized before, the model
will take care of it
"""
# Convert x into the right format
x = utils.check_data(x)
# Scaling the data
if self.auto_scaler:
if x.ndim == 1:
x = self.scaler.transform(x.reshape(1, -1))
elif x.ndim == 2:
x = self.scaler.transform(x)
else:
# Ensuring x has 2 dimensions
if x.ndim == 1:
x = np.reshape(x, (1, -1))
# Transforming into pytorch objects
x = torch.cuda.FloatTensor(x)
# Calculating risk_score
score = self.model(x).data.cpu().numpy().flatten()
if not use_log:
score = np.exp(score)
return score
def predict_cdf(self, x, t=None, **kwargs):
""" Predicts the cumulative density function F(t, x)
Parameters
----------
* `x` : **array-like** *shape=(n_samples, n_features)* --
array-like representing the datapoints.
x should not be standardized before, the model
will take care of it
* `t`: **double** *(default=None)* --
time at which the prediction should be performed.
If None, then return the function for all available t.
Returns
-------
* `cdf`: **numpy.ndarray** --
array-like representing the prediction of the cumulative
density function
"""
# Checking if the data has the right format
x = utils.check_data(x)
# Calculating survival and cdf
survival = self.predict_survival(x, t, **kwargs)
cdf = 1. - survival
return cdf
def bootstrap_concordance_index_chunk(model, X, T, E, include_ties = True, additional_results = False, n_iterations = 1000, n_size = 1000, **kwargs):
stats = list()
risk = model.predict_risk_chunk(X, **kwargs)
risk, T, E = utils.check_data(risk, T, E)
for i in range(n_iterations):
tempR, tempT, tempE = resample(risk, T, E, n_samples=n_size)
order = np.argsort(-tempT)
tempR = tempR[order]
tempT = tempT[order]
tempE = tempE[order]
# Calculating th c-index
results = _concordance_index(tempR, tempT, tempE, include_ties)
stats.append(results[0])
if i % 10 == 0:
print(i)
alpha = 0.95
p = ((1.0 - alpha) / 2.0) * 100
lower = max(0.0, np.percentile(stats, p))
p = (alpha + ((1.0 - alpha) / 2.0)) * 100
upper = min(1.0, np.percentile(stats, p))
print('%.1f confidence interval %.3f%% and %.3f%%' % (alpha * 100, lower, upper))
print(np.average(stats))
# confidence intervals
alpha = 0.95
p = ((1.0 - alpha) / 2.0) * 100
lower = max(0.0, np.percentile(stats, p))
p = (alpha + ((1.0 - alpha) / 2.0)) * 100
upper = min(1.0, np.percentile(stats, p))
print('%.1f confidence interval %.3f%% and %.3f%%' % (alpha * 100, lower, upper))
print(np.average(stats))
def predict(self, x, t=None):
""" Predicting the hazard, density and survival functions
Parameters:
----------
* `x` : **array-like** *shape=(n_samples, n_features)* --
array-like representing the datapoints.
x should not be standardized before, the model
will take care of it
* `t`: **double** *(default=None)* --
time at which the prediction should be performed.
If None, then return the function for all available t.
"""
# Convert x into the right format
x = utils.check_data(x)
# Scaling the data
if self.auto_scaler:
if x.ndim == 1:
x = self.scaler.transform(x.reshape(1, -1))
elif x.ndim == 2:
x = self.scaler.transform(x)
else:
# Ensuring x has 2 dimensions
if x.ndim == 1:
x = np.reshape(x, (1, -1))
# Transforming into pytorch objects
x = torch.FloatTensor(x)
# Predicting using linear/nonlinear function
score_torch = self.model(x)
score = score_torch.data.numpy()
# Cretaing the time triangles
Triangle1 = np.tri(self.num_times, self.num_times + 1)
Triangle2 = np.tri(self.num_times + 1, self.num_times + 1)
# Calculating the score, density, hazard and Survival
phi = np.exp(np.dot(score, Triangle1))
div = np.repeat(np.sum(phi, 1).reshape(-1, 1), phi.shape[1], axis=1)
density = (phi / div)
Survival = np.dot(density, Triangle2)
hazard = density[:, :-1] / Survival[:, 1:]
# Returning the full functions of just one time point
if t is None:
return hazard, density, Survival
else:
min_abs_value = [
abs(a_j_1 - t) for (a_j_1, a_j) in self.time_buckets
]
index = np.argmin(min_abs_value)
return hazard[:, index], density[:, index], Survival[:, index]
def predict_risk(self, x, use_log=False):
"""
Predicting the risk score functions
Arguments:
* x: pd.Dataframe or np.ndarray or list
x is the testing dataset containing the features
x should not be standardized before, the model
will take care of it
"""
# Convert x into the right format
x = utils.check_data(x)
if isinstance(x, list):
# Scaling data
if self.auto_scaler:
for index, X in enumerate(x):
if X.ndim == 1:
X = self.scaler.transform(X.reshape(1, -1))
elif X.ndim == 2:
X = self.scaler.transform(X)
x[index] = X
else:
for index, X in enumerate(x):
# Ensuring x has 2 dimensions
if X.ndim == 1:
X = np.reshape(X, (1, -1))
x[index] = X
else:
# Scaling data
if self.auto_scaler:
x = self.scaler.fit_transform(x)
# Transform into torch.Tensor
if isinstance(x, list):
for j, input_ in enumerate(x):
x[j] = torch.FloatTensor(input_)
if torch.cuda.is_available():
x[j] = x[j].cuda()
else:
x = torch.FloatTensor(x)
if torch.cuda.is_available():
x = x.cuda()
# Transforming into pytorch objects
#x = X_original
# Calculating risk_score
score = self.model(x).cpu().data.numpy().flatten()
if not use_log:
score = np.exp(score)
return score
def predict_risk(self, X, num_threads=-1):
# Checking if the data has the right format
X = utils.check_data(X)
if X.ndim == 1:
X = X.reshape(1, -1)
T = np.array([1.] * X.shape[0])
E = np.array([1.] * X.shape[0])
input_data = np.c_[T, E, X]
# Loading the attributes of the model
self.load_properties()
# Computing risk
risk = self.model.predict_risk(input_data, num_threads)
return np.array(risk)
def predict(self, x, t=None):
"""
Predicting the hazard, density and survival functions
Parameters:
-----------
* x: pd.Dataframe or np.ndarray or list
x is the testing dataset containing the features
x should not be standardized before, the model
will take care of it
* t: float (default=None)
Time at which hazard, density and survival functions
should be calculated. If None, the method returns
the functions for all times t.
"""
# Convert x into the right format
x = utils.check_data(x)
# Scaling the dataset
if x.ndim == 1:
x = self.scaler.transform(x.reshape(1, -1))
elif x.ndim == 2:
x = self.scaler.transform(x)
else:
# Ensuring x has 2 dimensions
if x.ndim == 1:
x = np.reshape(x, (1, -1))
# Calculating risk_score, hazard, density and survival
BX = self.risk_function(x)
hazard = self.hazard_function(self.times, BX.reshape(-1, 1))
survival = self.survival_function(self.times, BX.reshape(-1, 1))
density = (hazard * survival)
if t is None:
return hazard, density, survival
else:
min_abs_value = [
abs(a_j_1 - t) for (a_j_1, a_j) in self.time_buckets
]
index = np.argmin(min_abs_value)
return hazard[:, index], density[:, index], survival[:, index]
def predict_risk(self, x, use_log=False):
""" Predicts the Risk Score
Parameter
----------
* `x`, np.ndarray
array-like representing the datapoints
* `use_log`: bool - (default=False)
Applies the log function to the risk values
Returns
-------
* `risk_score`, np.ndarray
array-like representing the prediction of Risk Score function
"""
# Ensuring that the C++ model has the fitted parameters
self.load_properties()
# Convert x into the right format
x = utils.check_data(x)
# Scaling the dataset
if x.ndim == 1:
if self.with_bias:
x = np.r_[x, 1.]
x = self.scaler.transform(x.reshape(1, -1))
elif x.ndim == 2:
n = x.shape[0]
if self.with_bias:
x = np.c_[x, [1.] * n]
x = self.scaler.transform(x)
# Calculating prdiction
risk = np.exp(self.model.get_score(x))
if use_log:
return np.log(risk)
else:
return risk
def predict_chunk(self, X, t=1, num_threads=-1):
# Checking if the data has the right format
X = utils.check_data(X)
if X.ndim == 1:
X = X.reshape(1, -1)
T = np.array([1.] * X.shape[0])
E = np.array([1.] * X.shape[0])
input_data = np.c_[T, E, X]
n_rows, n_cols = input_data.shape
nparts = round(n_rows / 1000)
split_array = np.array_split(input_data, nparts)
survival = np.empty(n_rows)
self.load_properties()
min_index = [abs(a_j_1 - t) for (a_j_1, a_j) in self.time_buckets]
index = np.argmin(min_index)
survIndex = 0
for s in split_array:
survival[survIndex:(survIndex + np.size(s, axis=0))] = np.array(
self.model.predict_survival(s, num_threads))[:, index]
survIndex += np.size(s, axis=0)
#print(survIndex)
return survival
def fit(self,
X,
T,
E,
init_method='glorot_uniform',
optimizer='adam',
lr=1e-4,
num_epochs=1000,
l2_reg=1e-2,
verbose=True,
is_min_time_zero=True,
extra_pct_time=0.1):
"""
Fit the estimator based on the given parameters.
Parameters:
-----------
* `X` : **array-like**, *shape=(n_samples, n_features)* --
The input samples.
* `T` : **array-like** --
The target values describing when the event of interest or censoring
occurred.
* `E` : **array-like** --
The values that indicate if the event of interest occurred i.e.:
E[i]=1 corresponds to an event, and E[i] = 0 means censoring,
for all i.
* `init_method` : **str** *(default = 'glorot_uniform')* --
Initialization method to use. Here are the possible options:
* `glorot_uniform`: Glorot/Xavier uniform initializer
* `he_uniform`: He uniform variance scaling initializer
* `uniform`: Initializing tensors with uniform (-1, 1) distribution
* `glorot_normal`: Glorot normal initializer,
* `he_normal`: He normal initializer.
* `normal`: Initializing tensors with standard normal distribution
* `ones`: Initializing tensors to 1
* `zeros`: Initializing tensors to 0
* `orthogonal`: Initializing tensors with a orthogonal matrix,
* `optimizer`: **str** *(default = 'adam')* --
iterative method for optimizing a differentiable objective function.
Here are the possible options:
- `adadelta`
- `adagrad`
- `adam`
- `adamax`
- `rmsprop`
- `sparseadam`
- `sgd`
* `lr`: **float** *(default=1e-4)* --
learning rate used in the optimization
* `num_epochs`: **int** *(default=1000)* --
The number of iterations in the optimization
* `l2_reg`: **float** *(default=1e-4)* --
L2 regularization parameter for the model coefficients
* `verbose`: **bool** *(default=True)* --
Whether or not producing detailed logging about the modeling
* `extra_pct_time`: **float** *(default=0.1)* --
Providing an extra fraction of time in the time axis
* `is_min_time_zero`: **bool** *(default=True)* --
Whether the the time axis starts at 0
Returns:
--------
* self : object
Example:
--------
#### 1 - Importing packages
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
from sklearn.model_selection import train_test_split
from pysurvival.models.simulations import SimulationModel
from pysurvival.models.parametric import GompertzModel
from pysurvival.utils.metrics import concordance_index
from pysurvival.utils.display import integrated_brier_score
#%matplotlib inline # To use with Jupyter notebooks
#### 2 - Generating the dataset from a Gompertz parametric model
# Initializing the simulation model
sim = SimulationModel( survival_distribution = 'Gompertz',
risk_type = 'linear',
censored_parameter = 10.0,
alpha = .01, beta = 3.0 )
# Generating N random samples
N = 1000
dataset = sim.generate_data(num_samples = N, num_features = 3)
# Showing a few data-points
time_column = 'time'
event_column = 'event'
dataset.head(2)
#### 3 - Creating the modeling dataset
# Defining the features
features = sim.features
# Building training and testing sets #
index_train, index_test = train_test_split( range(N), test_size = 0.2)
data_train = dataset.loc[index_train].reset_index( drop = True )
data_test = dataset.loc[index_test].reset_index( drop = True )
# Creating the X, T and E input
X_train, X_test = data_train[features], data_test[features]
T_train, T_test = data_train['time'].values, data_test['time'].values
E_train, E_test = data_train['event'].values, data_test['event'].values
#### 4 - Creating an instance of the Gompertz model and fitting the data
# Building the model
gomp_model = GompertzModel()
gomp_model.fit(X_train, T_train, E_train, lr=1e-2, init_method='zeros',
optimizer ='adam', l2_reg = 1e-3, num_epochs=2000)
#### 5 - Cross Validation / Model Performances
c_index = concordance_index(gomp_model, X_test, T_test, E_test) #0.8
print('C-index: {:.2f}'.format(c_index))
ibs = integrated_brier_score(gomp_model, X_test, T_test, E_test,
t_max=30, figure_size=(20, 6.5) )
"""
# Checking data format (i.e.: transforming into numpy array)
X, T, E = utils.check_data(X, T, E)
T = np.maximum(T, 1e-6)
self.get_times(T, is_min_time_zero, extra_pct_time)
# Extracting data parameters
nb_units, self.num_vars = X.shape
input_shape = self.num_vars
# Scaling data
if self.auto_scaler:
X = self.scaler.fit_transform(X)
# Does the model need a parameter called Beta
is_beta_used = True
init_alpha = 1.
if self.name == 'ExponentialModel':
is_beta_used = False
if self.name == 'GompertzModel':
init_alpha = 1000.
# Initializing the model
model = nn.ParametricNet(input_shape, init_method, init_alpha,
is_beta_used)
# Trasnforming the inputs into tensors
X = torch.FloatTensor(X)
T = torch.FloatTensor(T.reshape(-1, 1))
E = torch.FloatTensor(E.reshape(-1, 1))
# Performing order 1 optimization
model, loss_values = opt.optimize(self.loss_function,
model,
optimizer,
lr,
num_epochs,
verbose,
X=X,
T=T,
E=E,
l2_reg=l2_reg)
# Saving attributes
self.model = model.eval()
self.loss_values = loss_values
# Calculating the AIC
self.aic = 2 * self.loss_values[-1]
self.aic -= 2 * (self.num_vars + 1 + is_beta_used * 1. - 1)
return self
def create_risk_groups(model,
X,
use_log=True,
num_bins=50,
figure_size=(20, 8),
**kwargs):
"""
Computing and displaying the histogram of the risk scores of the given
model and test set X. If it is provided args, it will assign a color coding
to the scores that are below and above the given thresholds.
Parameters:
-----------
* model : Pysurvival object
Pysurvival model
* X : array-like, shape=(n_samples, n_features)
The input samples.
* use_log: boolean (default=True)
Whether applying the log function to the risk score
* num_bins: int (default=50)
The number of equal-width bins that will constitute the histogram
* figure_size: tuple of double (default= (16, 6))
width, height in inches representing the size of the chart
* kwargs: dict (optional)
kwargs = low_risk = {'lower_bound': 0, 'upper_bound': 20, 'color': 'red'},
high_risk = {'lower_bound': 20, 'upper_bound': 120, 'color': 'blue'}
that define the risk group
"""
# Ensuring that the input data has the right format
X = utils.check_data(X)
# Computing the risk scores
risk = model.predict_risk(X)
if use_log:
risk = np.log(risk)
# Displaying simple histogram
if len(kwargs) == 0:
# Initializing the chart
fig, ax1 = plt.subplots(figsize=figure_size)
risk_groups = None
# Applying any color coding
else:
# Initializing the results
risk_groups = {}
# Initializing the chart
fig, ((ax1, ax2)) = plt.subplots(1, 2, figsize=figure_size)
# Displaying simple histogram with risk groups
nums_per_bins, bins, patches = ax2.hist(risk, bins=num_bins)
ax2.set_title('Risk groups with colors', fontsize=15)
# Number of group definitions
num_group_def = len(kwargs.values())
# Extracting the bounds values
bounds = {}
colors_ = {}
indexes = {}
group_names = []
handles = []
# we need to check that the boundaries match the bins
is_not_valid = 0
for group_name, group_def in kwargs.items():
# by ensuring that the bounds are not outside
# the bins values
min_bin, max_bin = min(bins), max(bins)
if (group_def['lower_bound'] < min_bin and \
group_def['upper_bound'] < min_bin) or \
(group_def['lower_bound'] > max_bin and \
group_def['upper_bound'] > max_bin):
is_not_valid += 1
# Extracting the bounds
bounds[group_name] = (group_def['lower_bound'],
group_def['upper_bound'])
# Extracting the colors
colors_[group_name] = group_def['color']
# Creating index placeholders
indexes[group_name] = []
group_names.append(group_name)
color_indv = group_def['color']
handles.append(Rectangle((0, 0), 1, 1, color=color_indv, ec="k"))
if is_not_valid >= num_group_def:
error_msg = "The boundaries definitions {} do not match"
error_msg += ", the values of the risk scores."
error_msg = error_msg.format(list(bounds.values()))
raise ValueError(error_msg)
# Assigning each rectangle/bin to its group definition
# and color
colored_patches = []
bin_index = {}
for i, bin_, patch_ in zip(range(num_bins), bins, patches):
# Check if the bin belongs to this bound def
for grp_name, bounds_ in bounds.items():
if bounds_[0] <= bin_ < bounds_[-1]:
bin_index[i] = grp_name
# Extracting color
color_ = colors_[grp_name]
if color_ not in colors.CSS4_COLORS:
error_msg = '{} is not a valid color'
error_msg = error_msg.format(colors_[grp_name])
raise ValueError(error_msg)
patch_.set_facecolor(color_)
# Saving the rectangles
colored_patches.append(patch_)
# Assigning each sample to its group
risk_bins = np.minimum(np.digitize(risk, bins, True), num_bins - 1)
for i, r in enumerate(risk_bins):
# Extracting the right group_name
group_name = bin_index[r]
indexes[group_name].append(i)
# Displaying the original distribution
ax1.hist(risk, bins=num_bins, color='black', alpha=0.5)
ax1.set_title('Risk Score Distribution', fontsize=15)
# Show everything
plt.show()
# Returning results
if risk_groups is not None:
for group_name in group_names:
result = (colors_[group_name], indexes[group_name])
risk_groups[group_name] = result
return risk_groups
def fit(self,
X,
T,
E,
max_features='sqrt',
max_depth=5,
min_node_size=10,
num_threads=-1,
weights=None,
sample_size_pct=0.63,
alpha=0.5,
minprop=0.1,
num_random_splits=100,
importance_mode='impurity_corrected',
seed=None,
save_memory=False):
"""
Arguments:
---------
* X : array-like, shape=(n_samples, n_features)
The input samples.
* T : array-like, shape = [n_samples]
The target values describing when the event of interest or censoring
occurred
* E : array-like, shape = [n_samples]
The Event indicator array such that E = 1. if the event occurred
E = 0. if censoring occurred
* max_features : int, float or string, optional (default="all")
The number of features to consider when looking for the best split:
- If int, then consider `max_features` features at each split.
- If float, then `max_features` is a fraction and
`int(max_features * n_features)` features are considered at each
split.
- If "sqrt", then `max_features=sqrt(n_features)`
- If "log2", then `max_features=log2(n_features)`.
* min_node_size : int(default=10)
The minimum number of samples required to be at a leaf node
* num_threads: int (Default: -1)
The number of jobs to run in parallel for both fit and predict.
If -1, then the number of jobs is set to the number of cores.
* weights: array-like, shape = [n_samples] (default=None)
Weights for sampling of training observations.
Observations with larger weights will be selected with
higher probability in the bootstrap
* sample_size_pct: double (default = 0.63)
Percentage of original samples used in each tree building
* alpha: float
For "maxstat" splitrule: Significance threshold to allow splitting.
* minprop: float
For "maxstat" splitrule: Lower quantile of covariate
distribution to be considered for splitting.
* num_random_splits: int (default=100)
For "extratrees" splitrule, it is the Number of random splits
to consider for each candidate splitting variable.
* importance_mode: (default=impurity_corrected)
Variable importance mode. Here are the 2 options:
- `impurity` or `impurity_corrected`:
it's the unbiased heterogeneity reduction developed
by Sandri & Zuccolotto (2008)
- "permutation" it's unnormalized as recommended by Nicodemus et al.
- "normalized_permutation" it's normalized version of the
permutation importance computations by Breiman et al.
* `seed`: int (default=None) --
seed used by the random number generator. If None, the current
timestamp converted in UNIX is used.
* save_memory: bool (default=False) --
Use memory saving splitting mode. This will slow down the model
training. So, only set to `True` if you encounter memory problems.
Example:
--------
#### 1 - Importing packages
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
from sklearn.model_selection import train_test_split
from pysurvival.models.simulations import SimulationModel
from pysurvival.models.survival_forest import ConditionalSurvivalForestModel
from pysurvival.utils.metrics import concordance_index
from pysurvival.utils.display import integrated_brier_score
#%matplotlib inline # To use with Jupyter notebooks
#### 2 - Generating the dataset from a Exponential parametric model
# Initializing the simulation model
sim = SimulationModel( survival_distribution = 'exponential',
risk_type = 'linear',
censored_parameter = 1,
alpha = 3)
# Generating N random samples
N = 1000
dataset = sim.generate_data(num_samples = N, num_features=4)
# Showing a few data-points
dataset.head(2)
#### 3 - Creating the modeling dataset
# Defining the features
features = sim.features
# Building training and testing sets #
index_train, index_test = train_test_split( range(N), test_size = 0.2)
data_train = dataset.loc[index_train].reset_index( drop = True )
data_test = dataset.loc[index_test].reset_index( drop = True )
# Creating the X, T and E input
X_train, X_test = data_train[features], data_test[features]
T_train, T_test = data_train['time'].values, data_test['time'].values
E_train, E_test = data_train['event'].values, data_test['event'].values
#### 4 - Creating an instance of the model and fitting the data.
# Building the model
csf = ConditionalSurvivalForestModel(num_trees=200)
csf.fit(X_train, T_train, E_train,
max_features="sqrt", max_depth=5, min_node_size=20,
alpha = 0.05, minprop=0.1)
#### 5 - Cross Validation / Model Performances
c_index = concordance_index(l_mtlr, X_test, T_test, E_test) #0.81
print('C-index: {:.2f}'.format(c_index))
ibs = integrated_brier_score(l_mtlr, X_test, T_test, E_test, t_max=30,
figure_size=(20, 6.5) )
"""
# Collecting features names
N, self.num_variables = X.shape
if isinstance(X, pd.DataFrame):
features = X.columns.tolist()
else:
features = ['x_{}'.format(i) for i in range(self.num_variables)]
all_data_features = ["time", "event"] + features
# Transforming the strings into bytes
all_data_features = utils.as_bytes(all_data_features,
python_version=PYTHON_VERSION)
# Checking the format of the data
X, T, E = utils.check_data(X, T, E)
if X.ndim == 1:
X = X.reshape(1, -1)
T = T.reshape(1, -1)
E = E.reshape(1, -1)
input_data = np.c_[T, E, X]
# Number of trees
num_trees = self.num_trees
# Seed
if seed is None:
seed = 0
# sample_size_pct
if not isinstance(sample_size_pct, float):
error = "Error: Invalid value for sample_size_pct, "
error += "please provide a value that is > 0 and <= 1."
raise ValueError(error)
if (sample_size_pct <= 0 or sample_size_pct > 1):
error = "Error: Invalid value for sample_size_pct, "
error += "please provide a value that is > 0 and <= 1."
raise ValueError(error)
# Split Rule
if self.splitrule.lower() == 'logrank':
split_mode = 1
alpha = 0
minprop = 0
num_random_splits = 1
elif self.splitrule.lower() == "maxstat":
split_mode = 4
num_random_splits = 1
# Maxstat splitting
if not isinstance(alpha, float):
error = "Error: Invalid value for alpha, "
error += "please provide a value that is > 0 and < 1."
raise ValueError(error)
if (alpha <= 0 or alpha >= 1):
error = "Error: Invalid value for alpha, "
error += "please provide a value between 0 and 1."
raise ValueError(error)
if not isinstance(minprop, float):
error = "Error: Invalid value for minprop, "
error += "please provide a value between 0 and 0.5"
raise ValueError(error)
if (minprop < 0 or minprop > 0.5):
error = "Error: Invalid value for minprop, "
error += "please provide a value between 0 and 0.5"
raise ValueError(error)
elif self.splitrule.lower() == 'extratrees':
split_mode = 5
alpha = 0
minprop = 0
# Number of variables to possibly split at in each node
self.max_features = max_features
if isinstance(self.max_features, str):
if self.max_features.lower() == 'sqrt':
num_variables_to_use = int(np.sqrt(self.num_variables))
elif 'log' in self.max_features.lower():
num_variables_to_use = int(np.log(self.num_variables))
elif self.max_features.lower() == 'all':
num_variables_to_use = self.num_variables
else:
raise ValueError("Unknown max features option")
elif isinstance(self.max_features, float) or \
isinstance(self.max_features, int):
if 0 < self.max_features < 1:
num_variables_to_use = int(self.num_variables *
self.max_features)
elif self.max_features >= 1:
num_variables_to_use = min(self.num_variables,
self.max_features)
if self.max_features > self.num_variables:
msg = "max features value is greater than the number of "
msg += "variables ({num_variables}) of the input X. "
msg += "So it was set to {num_variables}."
msg = msg.format(num_variables=self.num_variables)
warnings.warn(msg, UserWarning)
elif self.max_features <= 0:
raise ValueError("max features is a positive value")
else:
raise ValueError("Unknown max features option")
# Defining importance mode
if 'permutation' in importance_mode.lower():
if 'scaled' in importance_mode.lower() or \
'normalized' in importance_mode.lower():
importance_mode = 2
else:
importance_mode = 3
elif 'impurity' in importance_mode.lower():
importance_mode = 5
else:
error = "{} is not a valid importance mode".format(importance_mode)
raise ValueError(error)
# Weights
if weights is None:
case_weights = [1. / N] * N
else:
case_weights = utils.check_data(weights)
if abs(sum(case_weights) - 1.) >= 1e-4:
raise Exception(
"The sum of the weights needs to be equal to 1.")
if len(case_weights) != N:
raise Exception("weights length needs to be {} ".format(N))
# Fitting the model using the C++ object
verbose = True
self.model.fit(input_data, all_data_features, case_weights, num_trees,
num_variables_to_use, min_node_size, max_depth, alpha,
minprop, num_random_splits, sample_size_pct,
importance_mode, split_mode, verbose, seed, num_threads,
save_memory)
# Saving the attributes
self.save_properties()
self.get_time_buckets()
# Extracting the Variable Importance
self.variable_importance = {}
for i, value in enumerate(self.variable_importance_):
self.variable_importance[features[i]] = value
# Saving the importance in a dataframe
self.variable_importance_table = pd.DataFrame(
data={'feature': list(self.variable_importance.keys()),
'importance': list(self.variable_importance.values())
},
columns=['feature', 'importance']).\
sort_values('importance', ascending=0).reset_index(drop=True)
importance = self.variable_importance_table['importance'].values
importance = np.maximum(importance, 0.)
sum_imp = sum(importance) * 1.
self.variable_importance_table['pct_importance'] = importance / sum_imp
return self
def generate_data(self,
num_samples=100,
num_features=3,
feature_weights=None):
"""
Generating a dataset of simulated survival times from a given
distribution through the hazard function using the Cox model
Parameters:
-----------
* `num_samples`: **int** *(default=100)* --
Number of samples to generate
* `num_features`: **int** *(default=3)* --
Number of features to generate
* `feature_weights`: **array-like** *(default=None)* --
list of the coefficients of the underlying Cox-Model.
The features linked to each coefficient are generated
from random distribution from the following list:
* binomial
* chisquare
* exponential
* gamma
* normal
* uniform
* laplace
If None then feature_weights = [1.]*num_features
Returns:
--------
* dataset: pandas.DataFrame
dataset of simulated survival times, event status and features
Example:
--------
from pysurvival.models.simulations import SimulationModel
# Initializing the simulation model
sim = SimulationModel( survival_distribution = 'gompertz',
risk_type = 'linear',
censored_parameter = 5.0,
alpha = 0.01,
beta = 5., )
# Generating N Random samples
N = 1000
dataset = sim.generate_data(num_samples = N, num_features=5)
# Showing a few data-points
dataset.head()
"""
# Data parameters
self.num_variables = num_features
if feature_weights is None:
self.feature_weights = [1.] * self.num_variables
feature_weights = self.feature_weights
else:
feature_weights = utils.check_data(feature_weights)
if num_features != len(feature_weights):
error = "The length of feature_weights ({}) "
error += "and num_features ({}) are not the same."
error = error.format(len(feature_weights), num_features)
raise ValueError(error)
self.feature_weights = feature_weights
# Generating random features
# Creating the features
X = np.zeros((num_samples, self.num_variables))
columns = []
for i in range(self.num_variables):
key, X[:, i] = self.random_data(num_samples)
columns.append('x_' + str(i + 1))
X_std = self.scaler.fit_transform(X)
BX = self.risk_function(X_std)
# Building the survival times
T = self.time_function(BX)
C = np.random.normal(loc=self.censored_parameter,
scale=5,
size=num_samples)
C = np.maximum(C, 0.)
time = np.minimum(T, C)
E = 1. * (T == time)
# Building dataset
self.features = columns
self.dataset = pd.DataFrame(data=np.c_[X, time, E],
columns=columns + ['time', 'event'])
# Building the time axis and time buckets
self.times = np.linspace(0., max(self.dataset['time']), self.bins)
self.get_time_buckets()
# Building baseline functions
self.baseline_hazard = self.hazard_function(self.times, 0)
self.baseline_survival = self.survival_function(self.times, 0)
# Printing summary message
message_to_print = "Number of data-points: {} - Number of events: {}"
print(message_to_print.format(num_samples, sum(E)))
return self.dataset
def compare_to_actual(model,
X,
T,
E,
times=None,
is_at_risk=False,
figure_size=(16, 6),
metrics=['rmse', 'mean', 'median'],
**kwargs):
"""
Comparing the actual and predicted number of units at risk and units
experiencing an event at each time t.
Parameters:
-----------
* model : pysurvival model
The model that will be used for prediction
* X : array-like, shape=(n_samples, n_features)
The input samples.
* T : array-like, shape = [n_samples]
The target values describing when the event of interest or censoring
occured
* E : array-like, shape = [n_samples]
The Event indicator array such that E = 1. if the event occured
E = 0. if censoring occured
* times: array-like, (default=None)
A vector of timepoints.
* is_at_risk: bool (default=True)
Whether the function returns Expected number of units at risk
or the Expected number of units experiencing the events.
* figure_size: tuple of double (default= (16, 6))
width, height in inches representing the size of the chart
* metrics: str or list of str (default='all')
Indicates the performance metrics to compute:
- if None, then no metric is computed
- if str, then the metric is computed
- if list of str, then the metrics are computed
The available metrics are:
- RMSE: root mean squared error
- Mean Abs Error: mean absolute error
- Median Abs Error: median absolute error
Returns:
--------
* results: float or dict
Performance metrics
"""
# Initializing the Kaplan-Meier model
X, T, E = utils.check_data(X, T, E)
kmf = KaplanMeierModel()
kmf.fit(T, E)
# Creating actual vs predicted
N = T.shape[0]
# Defining the time axis
if times is None:
times = kmf.times
# Number of Expected number of units at risk
# or the Expected number of units experiencing the events
actual = []
actual_upper = []
actual_lower = []
predicted = []
if is_at_risk:
model_predicted = np.sum(model.predict_survival(X, **kwargs), 0)
for t in times:
min_index = [abs(a_j_1 - t) for (a_j_1, a_j) in model.time_buckets]
index = np.argmin(min_index)
actual.append(N * kmf.predict_survival(None, t))
actual_upper.append(N * kmf.predict_survival_upper(None, t))
actual_lower.append(N * kmf.predict_survival_lower(None, t))
predicted.append(model_predicted[index])
else:
model_predicted = np.sum(model.predict_density(X, **kwargs), 0)
for t in times:
min_index = [abs(a_j_1 - t) for (a_j_1, a_j) in model.time_buckets]
index = np.argmin(min_index)
actual.append(N * kmf.predict_density(None, t))
h = kmf.predict_hazard(None, t)
actual_upper.append(N * kmf.predict_survival_upper(None, t) * h)
actual_lower.append(N * kmf.predict_survival_lower(None, t) * h)
predicted.append(model_predicted[index])
# Computing the performance metrics
results = None
title = 'Actual vs Predicted'
if metrics is not None:
# RMSE
rmse = np.sqrt(mean_squared_error(actual, predicted))
# Median Abs Error
med_ae = median_absolute_error(actual, predicted)
# Mean Abs Error
mae = mean_absolute_error(actual, predicted)
if isinstance(metrics, str):
# RMSE
if 'rmse' in metrics.lower() or 'root' in metrics.lower():
results = rmse
title += "\n"
title += "RMSE = {:.3f}".format(rmse)
# Median Abs Error
elif 'median' in metrics.lower():
results = med_ae
title += "\n"
title += "Median Abs Error = {:.3f}".format(med_ae)
# Mean Abs Error
elif 'mean' in metrics.lower():
results = mae
title += "\n"
title += "Mean Abs Error = {:.3f}".format(mae)
else:
raise NotImplementedError(
'{} is not a valid metric function.'.format(metrics))
elif isinstance(metrics, list) or isinstance(metrics, numpy.ndarray):
results = {}
# RMSE
is_rmse = False
if any([('rmse' in m.lower() or 'root' in m.lower()) \
for m in metrics]):
is_rmse = True
results['root_mean_squared_error'] = rmse
title += "\n"
title += "RMSE = {:.3f}".format(rmse)
# Median Abs Error
is_med_ae = False
if any(['median' in m.lower() for m in metrics]):
is_med_ae = True
results['median_absolute_error'] = med_ae
title += "\n"
title += "Median Abs Error = {:.3f}".format(med_ae)
# Mean Abs Error
is_mae = False
if any(['mean' in m.lower() for m in metrics]):
is_mae = True
results['mean_absolute_error'] = mae
title += "\n"
title += "Mean Abs Error = {:.3f}".format(mae)
if all([not is_mae, not is_rmse, not is_med_ae]):
error = 'The provided metrics are not available.'
raise NotImplementedError(error)
# Plotting
fig, ax = plt.subplots(figsize=figure_size)
ax.plot(times, actual, color='red', label='Actual', alpha=0.8, lw=3)
ax.plot(times, predicted, color='blue', label='Predicted', alpha=0.8, lw=3)
plt.xlim(0, max(T))
# Filling the areas between the Survival and Confidence Intervals curves
plt.fill_between(times,
actual,
actual_lower,
label='Confidence Intervals - Lower',
color='red',
alpha=0.2)
plt.fill_between(times,
actual,
actual_upper,
label='Confidence Intervals - Upper',
color='red',
alpha=0.2)
# Finalizing the chart
plt.title(title, fontsize=15)
plt.legend(fontsize=15)
plt.show()
return results
def fit(self,
X,
T,
E,
init_method='glorot_normal',
lr=1e-2,
max_iter=100,
l2_reg=1e-2,
alpha=0.95,
tol=1e-3,
verbose=True):
"""
Fitting a proportional hazards regression model using
the Efron's approximation method to take into account tied times.
As the Hessian matrix of the log-likelihood can be
calculated without too much effort, the model parameters are
computed using the Newton_Raphson Optimization scheme:
W_new = W_old - lr*<Hessian^(-1), gradient>
Arguments:
---------
* `X` : **array-like**, *shape=(n_samples, n_features)* --
The input samples.
* `T` : **array-like** --
The target values describing when the event of interest or
censoring occurred.
* `E` : **array-like** --
The values that indicate if the event of interest occurred
i.e.: E[i]=1 corresponds to an event, and E[i] = 0 means censoring,
for all i.
* `init_method` : **str** *(default = 'glorot_uniform')* --
Initialization method to use. Here are the possible options:
* `glorot_uniform`: Glorot/Xavier uniform initializer
* `he_uniform`: He uniform variance scaling initializer
* `uniform`: Initializing tensors with uniform (-1, 1) distribution
* `glorot_normal`: Glorot normal initializer,
* `he_normal`: He normal initializer.
* `normal`: Initializing tensors with standard normal distribution
* `ones`: Initializing tensors to 1
* `zeros`: Initializing tensors to 0
* `orthogonal`: Initializing tensors with a orthogonal matrix,
* `lr`: **float** *(default=1e-4)* --
learning rate used in the optimization
* `max_iter`: **int** *(default=100)* --
The maximum number of iterations in the Newton optimization
* `l2_reg`: **float** *(default=1e-4)* --
L2 regularization parameter for the model coefficients
* `alpha`: **float** *(default=0.95)* --
Confidence interval
* `tol`: **float** *(default=1e-3)* --
Tolerance for stopping criteria
* `verbose`: **bool** *(default=True)* --
Whether or not producing detailed logging about the modeling
Example:
--------
#### 1 - Importing packages
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
from sklearn.model_selection import train_test_split
from pysurvival.models.simulations import SimulationModel
from pysurvival.models.semi_parametric import CoxPHModel
from pysurvival.utils.metrics import concordance_index
from pysurvival.utils.display import integrated_brier_score
#%pylab inline # To use with Jupyter notebooks
#### 2 - Generating the dataset from a Log-Logistic parametric model
# Initializing the simulation model
sim = SimulationModel( survival_distribution = 'log-logistic',
risk_type = 'linear',
censored_parameter = 10.1,
alpha = 0.1, beta=1.2 )
# Generating N random samples
N = 1000
dataset = sim.generate_data(num_samples = N, num_features = 3)
#### 3 - Creating the modeling dataset
# Defining the features
features = sim.features
# Building training and testing sets #
index_train, index_test = train_test_split( range(N), test_size = 0.2)
data_train = dataset.loc[index_train].reset_index( drop = True )
data_test = dataset.loc[index_test].reset_index( drop = True )
# Creating the X, T and E input
X_train, X_test = data_train[features], data_test[features]
T_train, T_test = data_train['time'].values, data_test['time'].values
E_train, E_test = data_train['event'].values, data_test['event'].values
#### 4 - Creating an instance of the Cox PH model and fitting the data.
# Building the model
coxph = CoxPHModel()
coxph.fit(X_train, T_train, E_train, lr=0.5, l2_reg=1e-2,
init_method='zeros')
#### 5 - Cross Validation / Model Performances
c_index = concordance_index(coxph, X_test, T_test, E_test) #0.92
print('C-index: {:.2f}'.format(c_index))
ibs = integrated_brier_score(coxph, X_test, T_test, E_test, t_max=10,
figure_size=(20, 6.5) )
References:
-----------
* https://en.wikipedia.org/wiki/Proportional_hazards_model#Tied_times
* Efron, Bradley (1974). "The Efficiency of Cox's Likelihood
Function for Censored Data". Journal of the American Statistical
Association. 72 (359): 557-565.
"""
# Collecting features names
N, self.num_vars = X.shape
if isinstance(X, pd.DataFrame):
self.variables = X.columns.tolist()
else:
self.variables = ['x_{}'.format(i) for i in range(self.num_vars)]
# Checking the format of the data
X, T, E = utils.check_data(X, T, E)
order = np.argsort(-T)
T = T[order]
E = E[order]
X = self.scaler.fit_transform(X[order, :])
self.std_scale = np.sqrt(self.scaler.var_)
# Initializing the model
self.model = _CoxPHModel()
# Creating the time axis
self.model.get_times(T, E)
# Initializing the parameters
W = np.zeros(self.num_vars)
W = opt.initialization(init_method, W, False).flatten()
W = W.astype(np.float64)
# Optimizing to find best parameters
epsilon = 1e-9
self.model.newton_optimization(X, T, E, W, lr, l2_reg, tol, epsilon,
max_iter, verbose)
# Saving the Cython attributes in the Python object
self.weights = np.array(self.model.W)
self.loss = self.model.loss
self.times = np.array(self.model.times)
self.gradient = np.array(self.model.gradient)
self.Hessian = np.array(self.model.Hessian)
self.inv_Hessian = np.array(self.model.inv_Hessian)
self.loss_values = np.array(self.model.loss_values)
self.grad2_values = np.array(self.model.grad2_values)
# Computing baseline functions
score = np.exp(np.dot(X, self.weights))
baselines = _baseline_functions(score, T, E)
# Saving the Cython attributes in the Python object
self.baseline_hazard = np.array(baselines[1])
self.baseline_survival = np.array(baselines[2])
del self.model
self.get_time_buckets()
# Calculating summary
self.get_summary(alpha)
return self
def fit(self,
X,
T,
E,
init_method='glorot_uniform',
optimizer='adam',
lr=1e-4,
num_epochs=1000,
dropout=0.2,
batch_normalization=False,
bn_and_dropout=False,
l2_reg=1e-5,
verbose=True):
"""
Fit the estimator based on the given parameters.
Parameters:
-----------
* `X` : **array-like**, *shape=(n_samples, n_features)* --
The input samples.
* `T` : **array-like** --
The target values describing when the event of interest or censoring
occurred.
* `E` : **array-like** --
The values that indicate if the event of interest occurred i.e.:
E[i]=1 corresponds to an event, and E[i] = 0 means censoring,
for all i.
* `init_method` : **str** *(default = 'glorot_uniform')* --
Initialization method to use. Here are the possible options:
* `glorot_uniform`: Glorot/Xavier uniform initializer
* `he_uniform`: He uniform variance scaling initializer
* `uniform`: Initializing tensors with uniform (-1, 1) distribution
* `glorot_normal`: Glorot normal initializer,
* `he_normal`: He normal initializer.
* `normal`: Initializing tensors with standard normal distribution
* `ones`: Initializing tensors to 1
* `zeros`: Initializing tensors to 0
* `orthogonal`: Initializing tensors with a orthogonal matrix,
* `optimizer`: **str** *(default = 'adam')* --
iterative method for optimizing a differentiable objective function.
Here are the possible options:
- `adadelta`
- `adagrad`
- `adam`
- `adamax`
- `rmsprop`
- `sparseadam`
- `sgd`
* `lr`: **float** *(default=1e-4)* --
learning rate used in the optimization
* `num_epochs`: **int** *(default=1000)* --
The number of iterations in the optimization
* `dropout`: **float** *(default=0.5)* --
Randomly sets a fraction rate of input units to 0
at each update during training time, which helps prevent overfitting.
* `l2_reg`: **float** *(default=1e-4)* --
L2 regularization parameter for the model coefficients
* `batch_normalization`: **bool** *(default=True)* --
Applying Batch Normalization or not
* `bn_and_dropout`: **bool** *(default=False)* --
Applying Batch Normalization and Dropout at the same time
* `verbose`: **bool** *(default=True)* --
Whether or not producing detailed logging about the modeling
Example:
--------
#### 1 - Importing packages
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
from sklearn.model_selection import train_test_split
from pysurvival.models.simulations import SimulationModel
from pysurvival.models.semi_parametric import NonLinearCoxPHModel
from pysurvival.utils.metrics import concordance_index
from pysurvival.utils.display import integrated_brier_score
#%matplotlib inline # To use with Jupyter notebooks
#### 2 - Generating the dataset from a nonlinear Weibull parametric model
# Initializing the simulation model
sim = SimulationModel( survival_distribution = 'weibull',
risk_type = 'Gaussian',
censored_parameter = 2.1,
alpha = 0.1, beta=3.2 )
# Generating N random samples
N = 1000
dataset = sim.generate_data(num_samples = N, num_features=3)
# Showing a few data-points
dataset.head(2)
#### 3 - Creating the modeling dataset
# Defining the features
features = sim.features
# Building training and testing sets #
index_train, index_test = train_test_split( range(N), test_size = 0.2)
data_train = dataset.loc[index_train].reset_index( drop = True )
data_test = dataset.loc[index_test].reset_index( drop = True )
# Creating the X, T and E input
X_train, X_test = data_train[features], data_test[features]
T_train, T_test = data_train['time'].values, data_test['time'].values
E_train, E_test = data_train['event'].values, data_test['event'].values
#### 4 - Creating an instance of the NonLinear CoxPH model and fitting
# the data.
# Defining the MLP structure. Here we will build a 1-hidden layer
# with 150 units and `BentIdentity` as its activation function
structure = [ {'activation': 'BentIdentity', 'num_units': 150}, ]
# Building the model
nonlinear_coxph = NonLinearCoxPHModel(structure=structure)
nonlinear_coxph.fit(X_train, T_train, E_train, lr=1e-3,
init_method='xav_uniform')
#### 5 - Cross Validation / Model Performances
c_index = concordance_index(nonlinear_coxph, X_test, T_test, E_test)
print('C-index: {:.2f}'.format(c_index))
ibs = integrated_brier_score(nonlinear_coxph, X_test, T_test, E_test,
t_max=10, figure_size=(20, 6.5) )
"""
# Checking data format (i.e.: transforming into numpy array)
X, T, E = utils.check_data(X, T, E)
# Extracting data parameters
N, self.num_vars = X.shape
input_shape = self.num_vars
# Scaling data
if self.auto_scaler:
X_original = self.scaler.fit_transform(X)
# Sorting X, T, E in descending order according to T
order = np.argsort(-T)
T = T[order]
E = E[order]
X_original = X_original[order, :]
self.times = np.unique(T[E.astype(bool)])
self.nb_times = len(self.times)
self.get_time_buckets()
# Initializing the model
model = nn.NeuralNet(input_shape, 1, self.structure, init_method,
dropout, batch_normalization, bn_and_dropout)
# Looping through the data to calculate the loss
X = torch.cuda.FloatTensor(X_original)
# Computing the Risk and Fail tensors
Risk, Fail = self.risk_fail_matrix(T, E)
Risk = torch.cuda.FloatTensor(Risk)
Fail = torch.cuda.FloatTensor(Fail)
# Computing Efron's matrices
Efron_coef, Efron_one, Efron_anti_one = self.efron_matrix()
Efron_coef = torch.cuda.FloatTensor(Efron_coef)
Efron_one = torch.cuda.FloatTensor(Efron_one)
Efron_anti_one = torch.cuda.FloatTensor(Efron_anti_one)
# Performing order 1 optimization
model, loss_values = opt.optimize(self.loss_function,
model,
optimizer,
lr,
num_epochs,
verbose,
X=X,
Risk=Risk,
Fail=Fail,
Efron_coef=Efron_coef,
Efron_one=Efron_one,
Efron_anti_one=Efron_anti_one,
l2_reg=l2_reg)
# Saving attributes
self.model = model.eval()
self.loss_values = loss_values
# Computing baseline functions
x = X_original
x = torch.cuda.FloatTensor(x)
# Calculating risk_score
score = np.exp(
self.model(torch.cuda.FloatTensor(x)).data.cpu().numpy().flatten())
baselines = _baseline_functions(score, T, E)
# Saving the Cython attributes in the Python object
self.times = np.array(baselines[0])
self.baseline_hazard = np.array(baselines[1])
self.baseline_survival = np.array(baselines[2])
return self
def fit(self, T, E, weights = None, alpha=0.95):
""" Fitting the model according to the provided data.
Parameters:
-----------
* `T` : **array-like** --
The target values describing when the event of interest or censoring
occurred.
* `E` : **array-like** --
The values that indicate if the event of interest occurred i.e.: E[i]=1
corresponds to an event, and E[i] = 0 symbols censoring, for all i.
* `weights` : **array-like** *(default = None)* --
Array of weights that are assigned to individual samples.
If not provided, then each sample is given a unit weight.
* `alpha`: **float** *(default = 0.05)* --
Significance level
Returns:
--------
* self : object
Example:
--------
# Importing modules
import numpy as np
from matplotlib import pyplot as plt
from pysurvival.utils.display import display_non_parametric
# %matplotlib inline #Uncomment when using Jupyter
# Generating random times and event indicators
T = np.round(np.abs(np.random.normal(10, 10, 1000)), 1)
E = np.random.binomial(1, 0.3, 1000)
# Initializing the KaplanMeierModel
from pysurvival.models.non_parametric import KaplanMeierModel
km_model = KaplanMeierModel()
# Fitting the model
km_model.fit(T, E, alpha=0.95)
# Displaying the survival function and confidence intervals
display_non_parametric(km_model)
# Initializing the SmoothKaplanMeierModel
from pysurvival.models.non_parametric import SmoothKaplanMeierModel
skm_model = SmoothKaplanMeierModel(bandwith=0.1, kernel='normal')
# Fitting the model
skm_model.fit(T, E)
# Displaying the survival function and confidence intervals
display_non_parametric(skm_model)
"""
# Checking the format of the data
T, E = utils.check_data(T, E)
# weighting
if weights is None:
weights = [1.]*T.shape[0]
# Confidence Intervals
z = stats.norm.ppf((1. - alpha) / 2.)
# Building the Kaplan-Meier model
survival = self.model.fit(T, E, weights, z)
if sum(survival) <= 0. :
mem_error = "The kernel matrix cannot fit in memory."
mem_error += "You should use a bigger bandwidth b"
raise MemoryError(mem_error)
# Saving all properties
self.save_properties()
# Generating the Survival table
if 'smooth' not in self.name.lower() :
self.get_survival_table()
def fit(self,
X,
T,
E,
with_bias=True,
init_method='glorot_normal',
lr=1e-2,
max_iter=100,
l2_reg=1e-4,
tol=1e-3,
verbose=True):
"""
Fitting a Survival Support Vector Machine model.
As the Hessian matrix of the log-likelihood can be
calculated without too much effort, the model parameters are
computed using the Newton_Raphson Optimization scheme:
W_new = W_old - lr*<Hessian^(-1), gradient>
Arguments:
---------
* `X` : array-like, shape=(n_samples, n_features)
The input samples.
* `T` : array-like, shape = [n_samples]
The target values describing when the event of interest or censoring
occurred
* `E` : array-like, shape = [n_samples]
The Event indicator array such that E = 1. if the event occurred
E = 0. if censoring occurred
* `with_bias`: bool (default=True)
Whether a bias should be added
* `init_method` : str (default = 'glorot_uniform')
Initialization method to use. Here are the possible options:
* 'glorot_uniform': Glorot/Xavier uniform initializer,
* 'he_uniform': He uniform variance scaling initializer
* 'uniform': Initializing tensors with uniform (-1, 1) distribution
* 'glorot_normal': Glorot normal initializer,
* 'he_normal': He normal initializer.
* 'normal': Initializing tensors with standard normal distribution
* 'ones': Initializing tensors to 1
* 'zeros': Initializing tensors to 0
* 'orthogonal': Initializing tensors with a orthogonal matrix,
* `lr`: float (default=1e-4)
learning rate used in the optimization
* `max_iter`: int (default=100)
The maximum number of iterations in the Newton optimization
* `l2_reg`: float (default=1e-4)
L2 regularization parameter for the model coefficients
* `alpha`: float (default=0.95)
Confidence interval
* `tol`: float (default=1e-3)
Tolerance for stopping criteria
* `verbose`: bool (default=True)
Whether or not producing detailed logging about the modeling
Example:
--------
#### 1 - Importing packages
import numpy as np
import pandas as pd
from pysurvival.models.svm import LinearSVMModel
from pysurvival.models.svm import KernelSVMModel
from pysurvival.models.simulations import SimulationModel
from pysurvival.utils.metrics import concordance_index
from sklearn.model_selection import train_test_split
from scipy.stats.stats import pearsonr
# %pylab inline # to use in jupyter notebooks
#### 2 - Generating the dataset from the parametric model
# Initializing the simulation model
sim = SimulationModel( survival_distribution = 'Log-Logistic',
risk_type = 'linear',
censored_parameter = 1.1,
alpha = 1.5, beta = 4)
# Generating N Random samples
N = 1000
dataset = sim.generate_data(num_samples = N, num_features = 4)
#### 3 - Splitting the dataset into training and testing sets
# Defining the features
features = sim.features
# Building training and testing sets #
index_train, index_test = train_test_split( range(N), test_size = 0.2)
data_train = dataset.loc[index_train].reset_index( drop = True )
data_test = dataset.loc[index_test].reset_index( drop = True )
# Creating the X, T and E input
X_train, X_test = data_train[features], data_test[features]
T_train, T_test = data_train['time'].values, data_test['time'].values
E_train, E_test = data_train['event'].values, data_test['event'].values
#### 4 - Creating an instance of the SVM model and fitting the data.
svm_model = LinearSVMModel()
svm_model = KernelSVMModel(kernel='Gaussian', scale=0.25)
svm_model.fit(X_train, T_train, E_train, init_method='he_uniform',
with_bias = True, lr = 0.5, tol = 1e-3, l2_reg = 1e-3)
#### 5 - Cross Validation / Model Performances
c_index = concordance_index(svm_model, X_test, T_test, E_test) #0.93
print('C-index: {:.2f}'.format(c_index))
#### 6 - Comparing the model predictions to Actual risk score
# Comparing risk scores
svm_risks = svm_model.predict_risk(X_test)
actual_risks = sim.predict_risk(X_test).flatten()
print("corr={:.4f}, p_value={:.5f}".format(*pearsonr(svm_risks,
actual_risks)))# corr=-0.9992, p_value=0.00000
"""
# Collecting features names
N, self.num_vars = X.shape
if isinstance(X, pd.DataFrame):
self.variables = X.columns.tolist()
else:
self.variables = ['x_{}'.format(i) for i in range(self.num_vars)]
# Adding a bias or not
self.with_bias = with_bias
if with_bias:
self.variables += ['intercept']
p = int(self.num_vars + 1. * with_bias)
# Checking the format of the data
X, T, E = utils.check_data(X, T, E)
if with_bias:
# Adding the intercept
X = np.c_[X, [1.] * N]
X = self.scaler.fit_transform(X)
# Initializing the parameters
if self.kernel_type == 0:
W = np.zeros((p, 1))
else:
W = np.zeros((N, 1))
W = opt.initialization(init_method, W, False).flatten()
W = W.astype(np.float64)
# Optimizing to find best parameters
self.model.newton_optimization(X, T, E, W, lr, l2_reg, tol, max_iter,
verbose)
self.save_properties()
return self
def concordance_index(model, X, T, E, include_ties = True,
additional_results=False, **kwargs):
"""
Computing the C-index based on *On The C-Statistics For Evaluating Overall
Adequacy Of Risk Prediction Procedures With Censored Survival Data* and
*Estimating the Concordance Probability in a Survival Analysis
with a Discrete Number of Risk Groups* and *Concordance for Survival
Time Data: Fixed and Time-Dependent Covariates and Possible Ties in
Predictor and Time
Similarly to the AUC, C-index = 1 corresponds to the best model
prediction, and C-index = 0.5 represents a random prediction.
Parameters:
-----------
* model : Pysurvival object
Pysurvival model
* X : array-like, shape=(n_samples, n_features)
The input samples.
* E : array-like, shape = [n_samples]
The Event indicator array such that E = 1. if the event occured
E = 0. if censoring occured
* include_ties: bool (default=True)
Specifies whether ties in risk score are included in calculations
* additional_results: bool (default=False)
Specifies whether only the c-index should be returned (False)
or if a dict of values should returned. the values are:
- c-index
- nb_pairs
- nb_concordant_pairs
Returns:
--------
* results: double or dict (if additional_results = True)
- results is the c-index (double) if additional_results = False
- results is dict if additional_results = True such that
results[0] = C-index;
results[1] = nb_pairs;
results[2] = nb_concordant_pairs;
Example:
--------
"""
# Checking the format of the data
risk = model.predict_risk(X, **kwargs)
risk, T, E = utils.check_data(risk, T, E)
# Ordering risk, T and E in descending order according to T
order = np.argsort(-T)
risk = risk[order]
T = T[order]
E = E[order]
# Calculating th c-index
results = _concordance_index(risk, T, E, include_ties)
if not additional_results:
return results[0]
else:
return results
def fit(self,
X,
T,
E,
init_method='glorot_uniform',
optimizer='adam',
lr=1e-4,
num_epochs=1000,
dropout=0.2,
l2_reg=1e-2,
l2_smooth=1e-2,
batch_normalization=False,
bn_and_dropout=False,
verbose=True,
extra_pct_time=0.1,
is_min_time_zero=True,
max_norm=1.0,
min_clamp_value=1e-8,
max_clamp_value=torch.finfo(torch.float32).max - 1):
""" Fit the estimator based on the given parameters.
Parameters:
-----------
* `X` : **array-like**, *shape=(n_samples, n_features)* --
The input samples.
* `T` : **array-like** --
The target values describing when the event of interest or censoring
occurred.
* `E` : **array-like** --
The values that indicate if the event of interest occurred i.e.:
E[i]=1 corresponds to an event, and E[i] = 0 means censoring,
for all i.
* `init_method` : **str** *(default = 'glorot_uniform')* --
Initialization method to use. Here are the possible options:
* `glorot_uniform`: Glorot/Xavier uniform initializer
* `he_uniform`: He uniform variance scaling initializer
* `uniform`: Initializing tensors with uniform (-1, 1) distribution
* `glorot_normal`: Glorot normal initializer,
* `he_normal`: He normal initializer.
* `normal`: Initializing tensors with standard normal distribution
* `ones`: Initializing tensors to 1
* `zeros`: Initializing tensors to 0
* `orthogonal`: Initializing tensors with a orthogonal matrix,
* `optimizer`: **str** *(default = 'adam')* --
iterative method for optimizing a differentiable objective function.
Here are the possible options:
- `adadelta`
- `adagrad`
- `adam`
- `adamax`
- `rmsprop`
- `sparseadam`
- `sgd`
* `lr`: **float** *(default=1e-4)* --
learning rate used in the optimization
* `num_epochs`: **int** *(default=1000)* --
The number of iterations in the optimization
* `dropout`: **float** *(default=0.5)* --
Randomly sets a fraction rate of input units to 0
at each update during training time, which helps prevent overfitting.
* `l2_reg`: **float** *(default=1e-4)* --
L2 regularization parameter for the model coefficients
* `l2_smooth`: **float** *(default=1e-4)* --
Second L2 regularizer that ensures the parameters vary smoothly
across consecutive time points.
* `batch_normalization`: **bool** *(default=True)* --
Applying Batch Normalization or not
* `bn_and_dropout`: **bool** *(default=False)* --
Applying Batch Normalization and Dropout at the same time
* `display_loss`: **bool** *(default=True)* --
Whether or not showing the loss function values at each update
* `verbose`: **bool** *(default=True)* --
Whether or not producing detailed logging about the modeling
* `extra_pct_time`: **float** *(default=0.1)* --
Providing an extra fraction of time in the time axis
* `is_min_time_zero`: **bool** *(default=True)* --
Whether the the time axis starts at 0
* `max_norm`: **float** *(default=1.0)* --
Max l2 norm for gradient clipping
**Returns:**
* self : object
Example:
--------
#### 1 - Importing packages
import numpy as np
import pandas as pd
from matplotlib import pyplot as plt
from sklearn.model_selection import train_test_split
from pysurvival.models.simulations import SimulationModel
from pysurvival.models.multi_task import LinearMultiTaskModel
from pysurvival.utils.metrics import concordance_index
#%matplotlib inline # To use with Jupyter notebooks
#### 2 - Generating the dataset from a Weibull parametric model
# Initializing the simulation model
sim = SimulationModel( survival_distribution = 'Weibull',
risk_type = 'linear',
censored_parameter = 10.0,
alpha = .01, beta = 3.0 )
# Generating N random samples
N = 1000
dataset = sim.generate_data(num_samples = N, num_features = 3)
# Showing a few data-points
time_column = 'time'
event_column = 'event'
dataset.head(2)
#### 3 - Creating the modeling dataset
# Defining the features
features = sim.features
# Building training and testing sets #
index_train, index_test = train_test_split( range(N), test_size = 0.2)
data_train = dataset.loc[index_train].reset_index( drop = True )
data_test = dataset.loc[index_test].reset_index( drop = True )
# Creating the X, T and E input
X_train, X_test = data_train[features], data_test[features]
T_train, T_test = data_train['time'].values, data_test['time'].values
E_train, E_test = data_train['event'].values, data_test['event'].values
#### 4 - Initializing a MTLR model and fitting the data.
# Building a Linear model
mtlr = LinearMultiTaskModel(bins=50)
mtlr.fit(X_train, T_train, E_train, lr=5e-3, init_method='orthogonal')
# Building a Neural MTLR
# structure = [ {'activation': 'Swish', 'num_units': 150}, ]
# mtlr = NeuralMultiTaskModel(structure=structure, bins=150)
# mtlr.fit(X_train, T_train, E_train, lr=5e-3, init_method='adam')
#### 5 - Cross Validation / Model Performances
c_index = concordance_index(mtlr, X_test, T_test, E_test) #0.95
print('C-index: {:.2f}'.format(c_index))
"""
# Checking data format (i.e.: transforming into numpy array)
X, T, E = utils.check_data(X, T, E)
input_shape = []
# Extracting data parameters
if isinstance(X, list):
nb_inputs = len(X)
for data in X:
nb_units, num_vars = data.shape
input_shape.append(num_vars)
# Scaling data
if self.auto_scaler:
for index, data in enumerate(X):
X[index] = self.scaler.fit_transform(data)
else:
nb_inputs = 1
nb_units, self.num_vars = X.shape
input_shape.append(self.num_vars)
# Scaling data
if self.auto_scaler:
X = self.scaler.fit_transform(X)
# Building the time axis, time buckets and output Y
X_cens, X_uncens, Y_cens, Y_uncens \
= self.compute_XY(X, T, E, is_min_time_zero, extra_pct_time)
# Initializing the model
model = nn.NeuralNet(input_shape, self.num_times, self.structure,
init_method, dropout, batch_normalization,
bn_and_dropout)
# Creating the Triangular matrix
Triangle = np.tri(self.num_times, self.num_times + 1, dtype=np.float32)
Triangle = torch.FloatTensor(Triangle)
if torch.cuda.is_available():
model = model.cuda()
Triangle = Triangle.cuda()
# Performing order 1 optimization
model, loss_values = opt.optimize(self.loss_function,
model,
optimizer,
lr,
num_epochs,
verbose,
X_cens=X_cens,
X_uncens=X_uncens,
Y_cens=Y_cens,
Y_uncens=Y_uncens,
Triangle=Triangle,
l2_reg=l2_reg,
l2_smooth=l2_smooth,
max_norm=max_norm,
min_clamp_value=min_clamp_value,
max_clamp_value=max_clamp_value)
# Saving attributes
self.model = model.eval()
self.loss_values = loss_values
return self
def brier_score(model, X, T, E, t_max=None, use_mean_point=True, **kwargs):
"""
Computing the Brier score at all times t such that t <= t_max;
it represents the average squared distances between
the observed survival status and the predicted
survival probability.
In the case of right censoring, it is necessary to adjust
the score by weighting the squared distances to
avoid bias. It can be achieved by using
the inverse probability of censoring weights method (IPCW),
(proposed by Graf et al. 1999; Gerds and Schumacher 2006)
by using the estimator of the conditional survival function
of the censoring times calculated using the Kaplan-Meier method,
such that :
BS(t) = 1/N*( W_1(t)*(Y_1(t) - S_1(t))^2 + ... +
W_N(t)*(Y_N(t) - S_N(t))^2)
In terms of benchmarks, a useful model will have a Brier score below
0.25. Indeed, it is easy to see that if for all i in [1,N],
if S(t, xi) = 0.5, then BS(t) = 0.25.
Parameters:
-----------
* model : Pysurvival object
Pysurvival model
* X : array-like, shape=(n_samples, n_features)
The input samples.
* T : array-like, shape = [n_samples]
The target values describing when the event of interest or censoring
occured
* E : array-like, shape = [n_samples]
The Event indicator array such that E = 1. if the event occured
E = 0. if censoring occured
* t_max: float
Maximal time for estimating the prediction error curves.
If missing the largest value of the response variable is used.
Returns:
--------
* (times, brier_scores):tuple of arrays
-times represents the time axis at which the brier scores were
computed
- brier_scores represents the values of the brier scores
Example:
--------
"""
# Checking the format of the data
T, E = utils.check_data(T, E)
# computing the Survival function
Survival = model.predict_survival(X, None, **kwargs)
# Extracting the time buckets
times = model.times
time_buckets = model.time_buckets
# Ordering Survival, T and E in descending order according to T
order = np.argsort(-T)
Survival = Survival[order, :]
T = T[order]
E = E[order]
if t_max is None or t_max <= 0.:
t_max = max(T)
# Calculating the brier scores at each t <= t_max
results = _brier_score(Survival, T, E, t_max, times, time_buckets,
use_mean_point)
times = results[0]
brier_scores = results[1]
return (times, brier_scores)
