Multivariate and regularized CoxPH models

[mhaist94]

Description of feature

Hi all,

thanks for taking your time digging through my requests:
Another important points (at least for clinicians, I will get to some more biology-focused features later) would be the implementation of multivariate CoxPH models that take into account multiple variables that might affect your time-to-event endpoint.
One of those (that is though limited in terms of the model fitting accuracy, and thus should not be employed with more than 5 variables at a time) is described in the SurvivalAnalysis (analyse_multivariate function, see https://cran.r-project.org/web/packages/survivalAnalysis/vignettes/multivariate.html) package in R or again in the survminer package in R (here its a coxph function if I recall correctly).
To fit more parameters into a multivariate model (which is rarely done), one usually employs regularized CoxPH models (for example LASSO regularized models, which is implmented in the glmnet package in R, see https://glmnet.stanford.edu/articles/glmnet.html). This is not a must-have - but given that the utility of ehrapy is particularly in the dissection of big heterogeneous datasets regularized models might be an idea worthwhile considering.

[eroell]

lifeline's CoxPHFitter, which ep.tl.cox_ph uses, does support multivariate CoxPH models already if I understood the request correctly; also, CoxPHFitterdoes allow for regularization, for example lasso, via the penalizer argument and the l1_ratio.

However, p.tl.cox_ph right now does not pass these arguments through; Extending the arguments passed to CoxPH here would make a lot of sense, and address this issue.

[eroell]

And, we could show the effect of this regularization also in the survival analysis notebook to highlight this option.