Output of Random Survival Forest

Question

I am trying to understand the Random Survival Forest output and how to use it to predict survival times of a cancer patients in an active drug study. I have stepped through this demo that uses breast cancer data from 686 individuals to train a survival time model and I don't quite understand the output of the trained model.

It seems that the Random Survival Forest is being trained on data where each row corresponds to a patient and the covariates/predictors (estrogen levels, tumor size, etc) are from the "endpoint" of the study: either when the patient passes or when the study ends.

The demo then uses the test data (which similarly has "endpoint" covariate measurements) to score the model. The output of rsf.predict_survival_function(X_test_sel, return_array=True) gives an array that represents the survival probability over time for each patient in X_test_sel. Is the correct interpretation of this array the survival probabilities for time t where t is after the measurement time given as input to the model or is t the time from the start of the study?

Additionally, wouldn't a more realistic test scenario be using covariate measurements of a patient at the start of the study (and not at the endpoint)? Can a model be trained on "endpoint" covariate measurements be used to actually predict the survival time of a patient whose future covariate measurements are not known (where only current measurements at the start or midpoint of the drug trial are given)?

Apologies in advance if these are very basic questions; I am new to the area.

Answer 1

You can't understand the output before you understand the inputs.

In this case you seem to have jumped to a conclusion that the covariates were measured at the endpoint simply because this is a survival analysis of breast cancer patients.

I read the description of the German Breast Cancer Study Group 2 (GBSG2) dataset in the scikit-survival documentation and looked up the references therein to learn that the study recruited patients who already had a breast cancer diagnosis. So the covariates / prognostic factors (age, tumor size, etc.) were recorded at the time of the primary diagnosis, i.e., before the subjects were enrolled in GBSG2.

Sauerbrei W et al. Randomized 2 x 2 trial evaluating hormonal treatment and the duration of chemotherapy in node-positive breast cancer patients: an update based on 10 years' follow-up. German Breast Cancer Study Group. J Clin Oncol. 2000 Jan;18(1):94-101. HTML version