Pycox

Release notes


Features

- Administrative Brier score and Binomial log-likelihood for evaluation of data sets with administrative censoring.

- `BCESurv` which is a method that disregards censoring and does not enforce monotone survival functions. It is meant to represent a set of binary classifiers that disregards censored observations.

- Improved `kkbox` data sets with administrative censoring times and more covariates.

- `sac_admin5` simulated data set with administrative censoring.

- More simulations studies with covariate dependent censoring times and administrative censoring.


Changes

- Updated to work with `torchtuples` v.0.2.0

- `CoxPH` now use a regular data set, instead of the durations sorted. The old method is renamed `CoxPHSorted` but will be removed.

Minor bug fixes and release to PyPI.

These note mainly focus on the changes to existing code and not new functionality.

Evaluation criteria

- Fixed wrong index in IPCW.
- `EvalSurv` now has a `steps` argument determining how the survival curve should behave between estimated times.
Previously set to 'pre', but now 'post' is default.
This will affect the concordance for the discrete-time methods the most. Set `ev.step = 'pre'` to obtain old results.
Or use some reasonable interpolation scheme.
- Moved `pycox.evaluation.utils` to `pycox.utils`.
- Replaced the binomial log-likelihood `mbll` with the negative binomial log-likelihood `nbll`. I.e. only the sign is different.

Models

- Replaced `predict_survival_function` with `predict_surv` and `predict_surv_df`.
- More stable version of CoxCC and CoxTime loss for single control.
- Restructured the locations of the Cox models.

Preprocessing

- Added quantiles discretization for methods.