Hi,

When creating a new metric type, we need to specify the associated prediction type, as defined in tune:::pred_type.

Currently, there is no metric type that triggers a type = "linear_pred" prediction.

However, several metrics rely on the linear predictor (also known as log-partial hazard or pronostic index), such as calibration slope, Royston & Sauerbrei $R_D^2$, Gönen & Heller’s K-index, etc. Technical details can be found in survival::royston or Royston & Altman (2013).

Would it be possible to implement a new ?_survival_metric class to handle linear_pred prediction, enabling users to define custom metrics that use this output?

Thanks for considering this feature request, and thanks a lot for your work on the tidymodels ecosystem!