Predict log partial hazard about pycox HOT 7 OPEN

Using the definition of the log partial hazard from here, for the DeepSurv, i.e., CoxPH model, you can simply call model.predict(x), as the log partial hazard is the output of the neural network.
The same is true for the CoxCC model.
For Cox-Time, this log partial hazard is a function of time, so I assume that is not what you're looking for?

from pycox.

We should probably add predict_partial_hazards and predict_partial_log_hazard this the the CoxPH and CoxCC models as is simplifies the usage. Also, they are a part of lifelines, so the user would expect them to be there.

from pycox.

Using the definition of the log partial hazard from here, for the DeepSurv, i.e., CoxPH model, you can simply call model.predict(x), as the log partial hazard is the output of the neural network.
The same is true for the CoxCC model.
For Cox-Time, this log partial hazard is a function of time, so I assume that is not what you're looking for?

I'm not well-versed in survival models but aren't partial hazard rates are dependent to time? We can predict with model.predict(X) but time vector is the first element of y tuple. I was trying to get some output from the model in order to calculate an oof concordance index, but this didn't seem right to me. How should I proceed?

from pycox.

@gunesevitan the idea of CoxPH (proportional hazards) is that the hazard rate is that the hazard of two individuals are represented by a scaling factor (which does not depend on time). So the hazard is defined as the (time-dependent) baseline hazard rate multiplied by this scaling factor (sometimes referred to as the partial hazard).
Therefore, the partial hazard rate of CoxPH should not be dependent on time.

I'm not sure I understand the second part of your question @gunesevitan. Could you try to explain again what you want to do?

from pycox.

@havakv So you mean that I can get hazard ratio from DeepHit using model.predict(x) as well?
Thank you so much in advance !

from pycox.

@kyuchoi No, the explanation above only works for proportional hazard models. Those models are constructed as the product of the baseline hazard and the partial hazard. For models that are not constructed this way (like deephit), I'm not sure the partial hazard is really obtainable. It would at least have to be dependent on time for some of them. I don't know it the partial hazard is that interesting to study or non-PH models at all, to be honest.

from pycox.

I'm sorry @kyuchoi, I thought you asked for the partial hazards. The answer is, however, very similar for the hazard ratios. You will not be able to get them directly for the DeepHit model. I'm not sure how to best study the hazard ratios for non-proportional hazards models, so I'm not really able to assist you here. However, if you find a way to get hazard ratios for non-proportional hazards model, I'd be happy to assist you with the implementation

from pycox.