Time-varying covariates for random survival forest

[danielwiegand]

I read that time-varying covariates are not supported by scikit-survival (#293 and #8). However, I am not sure if this applies only to Cox PH? When I have a dataset in the "long" format (one line per sample which is alive and per timestep where a sample "died"), is it somehow invalid to feed it for example into a random survival forest?

[sebp]

I believe you would need support for an entry time, because you want to express when the time slice started too. Currently, scikit-survival does not support this.

[danielwiegand]

Thank you, I guess I understand a bit better. I come from the survival stacking context - but the difference seems to be that classical survival approaches model survival time (and need start and stop of an interval) while for survival stacking, the target variable is whether a sample survived (boolean), given some covariables and survival time up to this point (the time variable may be one-hot encoded here). Time as covariable doesn't need to be an interval here as it's considered a point in time, while modeling the hazard / survival function looks at time intervals.

[sebp]

Both approaches solve the same problem: how to predict right-censored survival times. The difference is that Cox PH (and most traditional models) treat time as a continuous variable, whereas in survival stacking, time is considered to be discrete, thus resulting in a classification problem per discrete interval.

To model time-dependent features, you also transform the data into non-overlapping intervals. But to ensure only a single interval is considered for each sample/subject, an entry time is added.