Module auton_survival.metrics

Tools to compute metrics used to assess survival outcomes and survival model performance.

Browse git

Functions

def treatment_effect(metric, outcomes, treatment_indicator, weights=None, horizons=None, risk_levels=None, interpolate=True, weights_clip=0.01, n_bootstrap=None, size_bootstrap=1.0, random_seed=0)

Compute metrics for comparing population level survival outcomes across treatment arms.

Parameters

metric : str
 
The metric to evalute for comparing survival outcomes.
Options include:
- median
- tar
- hazard_ratio
- restricted_mean
- survival_at
outcomes : pd.DataFrame
A pandas dataframe with rows corresponding to individual samples and columns 'time' and 'event'.
treatment_indicator : np.array
Boolean numpy array of treatment indicators. True means individual was assigned a specific treatment.
weights : pd.Series, default=None
Treatment assignment propensity scores, \widehat{\mathbb{P}}(A|X=x) . If None, all weights are set to 0.5 . Default is None.
horizons : float or int or array of floats or ints, default=None
Time horizon(s) at which to compute the metric. Must be specified for metric 'restricted_mean' and 'survival_at'. For 'hazard_ratio' this is ignored.
risk_levels : float or array of floats
The risk level (0-1) at which to compare times between treatment arms. Must be specified for metric 'tar'. Ignored for other metrics.
interpolate : bool, default=True
Whether to interpolate the survival curves.
weights_clip : float
Weights below this value are clipped. This is to ensure IPTW estimation is numerically stable. Large weights can result in estimator with high variance.
n_bootstrap : int, default=None
The number of bootstrap samples to use. If None, bootrapping is not performed.
size_bootstrap : float, default=1.0
The fraction of the population to sample for each bootstrap sample.
random_seed : int, default=0
Controls the reproducibility random sampling for bootstrapping.

Returns

float or list: The metric value(s) for the specified metric.

Browse git
def survival_regression_metric(metric, outcomes, predictions, times, outcomes_train=None)

Compute metrics to assess survival model performance.

Parameters

metric : string
Measure used to assess the survival regression model performance. Options include: - brs : brier score - ibs : integrated brier score - auc: cumulative dynamic area under the curve - ctd : concordance index inverse probability of censoring weights (ipcw)
outcomes : pd.DataFrame
A pandas dataframe with rows corresponding to individual samples and columns 'time' and 'event' for evaluation data.
predictions : np.array
A numpy array of survival time predictions for the samples.
times : np.array
The time points at which to compute metric value(s).
outcomes_train : pd.DataFrame
A pandas dataframe with rows corresponding to individual samples and columns 'time' and 'event' for training data.

Returns

float: The metric value for the specified metric.

Browse git
def phenotype_purity(phenotypes_train, outcomes_train, phenotypes_test=None, outcomes_test=None, strategy='instantaneous', horizons=None, bootstrap=None)

Compute the brier score to assess survival model performance for phenotypes.

Parameters

phenotypes_train : np.array
A numpy array containing an array of integers that define subgroups for the train set.
outcomes_train : pd.DataFrame
A pandas dataframe with rows corresponding to individual samples and columns 'time' and 'event' for the train set.
phenotypes_test : np.array
A numpy array containing an array of integers that define subgroups for the test set.
outcomes_test : pd.DataFrame
A pandas dataframe with rows corresponding to individual samples and columns 'time' and 'event' for the test set.
strategy : string, default='instantaneous'
Options include: - instantaneous : Compute the brier score. - integrated : Compute the integrated brier score.
horizons : float or int or an array of floats or ints, default=None
Event horizon(s) at which to compute the metric
bootstrap : integer, default=None
The number of bootstrap iterations.

Returns

list:
Columns are metric values computed for each event horizon. If bootstrapping, rows are bootstrap results.
Browse git