cenreg.metric package

Submodules

cenreg.metric.cdf module

cenreg.metric.cdf.brier(dist, y: ndarray, y_bins: ndarray) → ndarray

Compute the (original) Brier score.

Parameters:

  • dist (predicted distribution)

  • y (Array of shape [batch_size])

  • y_bins (Array of shape [num_bins+1])

Returns:

loss

Return type:

Array of shape [batch_size]

cenreg.metric.cdf.km_calibration(dist, observed_times: ndarray, uncensored: ndarray | None = None, y_bins: ndarray | None = None, eps: float = 0.0001) → float

Compute KM-Calibration

Parameters:

  • dist (distribution object) – Prediction results.

  • observed_times (ndarray) – Observation time (event time or censored time) Array shape is [batch_size].

  • uncensored (ndarray) – Censored (False) or uncensored (True) Array shape is [batch_size].

  • y_bins (ndarray) – Bins for the prediction to be evaluated. Each element in observed_times must be STRICTLY smaller than y_bins[-1]

  • eps (float) – Small positive value for numerical stability

Returns:

KM-Calibration – A non-negative float number.

Return type:

float

cenreg.metric.cdf.negative_log_likelihood_interval(dist, lb: ndarray, ub: ndarray, proportional: bool = True, eps: float = 0.0001) → ndarray

Compute Negative log-likelihood for interval-censored data.

Parameters:

  • dist (predicted distribution) – Must be an instance of cenreg.distribution.cdf.CumulativeDist with linear interpolation.

  • lb (ndarray of shape [batch_size]) – lower bounds of the intervals

  • ub (ndarray of shape [batch_size]) – upper bounds of the intervals

  • proportional (bool) – whether to distribute the probability mass proportionally for censored data

  • eps (float) – small value to avoid numerical issues

Returns:

loss

Return type:

ndarray of shape [batch_size]

cenreg.metric.cdf.negative_log_likelihood_survival(dist, y: ndarray, uncensored: ndarray | None = None, proportional: bool = True, eps: float = 0.0001) → ndarray

Compute Negative log-likelihood for survival data.

Parameters:

  • dist (predicted distribution) – Must be an instance of cenreg.distribution.cdf.CumulativeDist with linear interpolation.

  • y (ndarray of shape [batch_size]) – observed times

  • uncensored (ndarray of shape [batch_size]) – boolean array indicating whether the event was observed (True) or censored (False)

  • proportional (bool) – whether to distribute the probability mass proportionally for censored data

  • eps (float) – small value to avoid numerical issues

Returns:

loss

Return type:

ndarray of shape [batch_size]

cenreg.metric.cdf.nll_sc(list_dist, observed_times: ndarray, events: ndarray, survival_copula, eps: float = 0.0001)

Compute Negative Log-Likelihood based on Survival Copula (NLL-SC).

cenreg.metric.cdf.ranked_probability_score(dist, y: ndarray, y_bins: ndarray) → ndarray

Compute the ranked probability score.

Parameters:

  • dist (predicted distribution)

  • y (Array of shape [batch_size])

  • y_bins (Array of shape [num_col])

Returns:

loss

Return type:

Array of shape [batch_size]

cenreg.metric.cjd module

cenreg.metric.cjd.brier(observed_times: ndarray, events: ndarray, num_risks: int, f_pred: ndarray, boundaries: ndarray) → float

Compute Brier score for censoree joint distribution.

Parameters:

  • observed_times (ndarray) – Observed times (event times or censoring times). Array shape is [batch_size].

  • events (ndarray) – Event index (0 usually corresponds to censoring) Array shape is [batch_size].

  • num_risks (int) – Number of risks.

  • f_pred (ndarray) – Estimated joint distribution. Array shape is [batch_size, num_bin*num_risks].

  • boundaries (ndarray) – Bin boundaries used to represent f_pred. boundaries[0] = 0.0 and boundaries[-1] = max_time. Array shape is [num_bin+1].

Returns:

Brier – Value of Brier score.

Return type:

float

cenreg.metric.cjd.kolmogorov_smirnov_calibration_error(observed_times: ndarray, events: ndarray, f_pred: ndarray, boundaries: ndarray) → float

Sum of Kolmogorov-Sminov calibration error.

Parameters:

  • observed_times (ndarray) – Observed times (event times or censoring times). Array shape is [batch_size].

  • events (ndarray) – Event index (0 usually corresponds to censoring) Array shape is [batch_size].

  • f_pred (ndarray) – Estimated joint distribution. Array shape is [batch_size, num_bin*num_risks].

  • boundaries (ndarray) – Bin boundaries used to represent f_pred. boundaries[0] = 0.0 and boundaries[-1] = max_time. Array shape is [num_bin+1].

Returns:

KS – Sum of Kolmogorov-Sminov calibration error.

Return type:

float

cenreg.metric.cjd.negative_loglikelihood(observed_times: ndarray, events: ndarray, f_pred: ndarray, boundaries: ndarray, epsilon: float = 0.0001) → ndarray

Compute negative log-likelihood (NLL).

Parameters:

  • observed_times (ndarray) – Observed times (event times or censoring times). Array shape is [batch_size].

  • events (ndarray) – Event index (0 usually corresponds to censoring) Array shape is [batch_size].

  • f_pred (ndarray) – Estimated joint distribution. Array shape is [batch_size, num_bin*num_risks].

  • boundaries (ndarray) – Bin boundaries used to represent f_pred. boundaries[0] = 0.0 and boundaries[-1] = max_time. Array shape is [num_bin+1].

  • epsilon (float) – Small value to avoid log(0).

Returns:

NLL – Per-sample negative log-likelihood values. Array shape is [batch_size].

Return type:

ndarray

cenreg.metric.cjd.ranked_probability_score(observed_times: ndarray, events: ndarray, num_risks: int, f_pred: ndarray, boundaries: ndarray) → float

Ranked probability score (RPS).

Parameters:

  • observed_times (ndarray) – Observed times (event times or censoring times). Array shape is [batch_size].

  • events (ndarray) – Event index (0 usually corresponds to censoring) Array shape is [batch_size].

  • num_risks (int) – Number of risks.

  • f_pred (ndarray) – Estimated joint distribution. Array shape is [batch_size, num_bin*num_risks].

  • boundaries (ndarray) – Bin boundaries used to represent f_pred. boundaries[0] = 0.0 and boundaries[-1] = max_time. Array shape is [num_bin+1].

Returns:

RPS – Value of ranked probability score.

Return type:

float

cenreg.metric.quantile module

cenreg.metric.quantile.d_calibration(dist, observed_times: ndarray, uncensored: ndarray, boundaries: ndarray | int | None = None) → float

Compute D-Calibration.

Parameters:

  • dist (distribution object) – Predicted distribution.

  • observed_times (ndarray (float) of shape [batch_size]) – Observation time (event time or censored time)

  • uncensored (ndarray (bool) of shape [batch_size]) – Indicator (censored (False) or uncensored (True))

  • boundaries (ndarray (float) of shape [num_bin+1] or int or None) – Bin boundaries used to compute D-calibration. boundaries[0] = 0.0 and boundaries[-1] = 1.0. If boundaries is int, then boundaries = np.linspace(0.0, 1.0, boundaries + 1). If boundaries is None, then boundaries = np.linspace(0.0, 1.0, 11).

Returns:

D-Calibration – Value of D-Calibration.

Return type:

float

cenreg.metric.quantile.ic_calibration(dist, lb: ndarray, ub: ndarray, p: float = 2.0, b: ndarray | None = None, EPS: float = 0.0001) → float

Interval-Censored Calibration (IC-Cal).

Parameters:

  • dist (distribution object) – Predicted distribution.

  • lb (ndarray (float) of shape [batch_size]) – Lower bound of the interval.

  • ub (ndarray (float) of shape [batch_size]) – Upper bound of the interval.

  • p (float) – Power for the calibration error. Only p=2 is implemented.

  • b (ndarray (float) or None) – Bin boundaries used to compute IC-Cal. If None, default boundaries are used.

  • EPS (float) – Small value to avoid division by zero.

Returns:

IC-Cal – Value of IC-Cal.

Return type:

float

Module contents