C-index score with predict & evaluation windows

[ziyit]

Dear Dr. Lee,

Please refer to "tutorial.ipynb" notebook, section "1. Import Dataset", the following pred_time and eval_time were defined.

# This must be changed depending on the datasets, prediction/evaliation times of interest pred_time = [52, 352, 552] # prediction time (in months) eval_time = [12, 36, 60, 120] # months evaluation time (for C-index and Brier-Score)

I would like to confirm the logic on providing these two windows. (1) pred_time outputs 52th, 352th, and 552th month risk for each patient (2) eval_time scores the risk of each patient at 12th, 36th, 60th, and 120th month (3) the C-index is computed as a 3-by-4 matrix, can you illustrate what does (pred_time == 52 & eval_time==12), the C-index imply?

Thank you so much!

[javahaa]

Hey,bro! did you have any idea about your issue? I have a question about "52"?Where did “52” come from？ Can you answer me?Thank you very much.

[ziyit]

Hi there, thank you for the followup. I suppose the "52" was a number selected by the author as a "semi-random" guess point between 36 and 60 to better evaluate the hazard (the late stage of eval time w/ ground truth observed). (The only way I can make sense of it is by spreading out the prediction window beyond eval time as you'd expected event(s) hazards/prob converge at the end (352 and 552 for example)).

If you would consider other DL method for survival, I'd check on recurrent deep survival machines by Auton Lab, which also cited Dr. Lee's work.

Regards.

[felipezeiser]

Hey guys,

I know this explanation is a bit late, but it might help you or others.

In the example, the time is given in weeks, which I discovered by looking at the original PBC2 dataset.

Therefore, the value of 52 corresponds to one year (approximately).

The pred_time vector defines which samples will be used to generate the risk of events in the eval_time interval. In other words, samples up to 52 weeks will be used to predict the risk of an event in the next 12 weeks, 36 weeks, 60 weeks, and 120 weeks.

In the next round, samples up to 156 weeks (approximately three years) are used, and the next 12, 36, and 120 weeks are predicted again.

Individuals who have already experienced an event up to pred_time are disregarded for the prediction. Therefore, the individual must still be alive at the instant of time analyzed in the pred_time vector.