search

havakv / pycox

Survival analysis with PyTorch
BSD 2-Clause "Simplified" License
803 stars 188 forks source link

Could you please tell me the pros and cons of your time-dependent C-index, compared to td AUC? #115

Open kyuchoi opened 2 years ago

kyuchoi commented 2 years ago

I want to use your timd-dependent C-index instead of time-dependent AUC.

However, I could not tell what is the reason to use time-dependent C-index instead of time-dependent AUC, which seems to be more widely used.

Could you please tell me the pros and cons of your time-dependent C-index, compared to td AUC?

Best,

havakv commented 2 years ago

Hi, the reasons for using the time-dependent C-index instead of AUC was a bit arbitrary. From my brief overview of the machine learning literature addressing survival issues, I found that the C index (Harrel's) was used a lot, so we just used the time dependente version by Antolini. Personally, I think the C index is more intuitive than the AUC, as it gives a single number that represents a probability estimate, while the AUC is a function of time. You could of course integrate the AUC(t) to get such a number, but then you are quickly in the territory of a C index again (you should be able to find some references for this).

Both metrics have their issues, in particular the assumptions they have on the censoring distribution, so you would probably be better off using both in your analysis if that is possible.