package-analysis | Open Source Package Analysis | Code Analyzer library

Amer. Thx for your reading of my tutorial!

As you mentioned that "The output returned from the predict function represents the f(x) component of the hazard function ( h(t|x)=lambda(t)*exp(f(x)) )", maybe we need to understand the hazard function, i.e. h(t|x).

Before this, please sure that you have the basic knowledge of survival analysis. if not, it's recommended to read the great post. I think the post would help you solve the questions.

Back to your questions:

• Exactly, we can get the hazard ratios of log scale by invoking the predict function. Therefore, the hazard ratio can be calculated by exp() .
• Sure! Relying on the values of hazard ratio, we can divide the population into low-risk and high-risk groups. Alternatively, you can use the median of hazard ratios as the cutoff value. I think the cutoff value should be derived from the training set, and then test in the test set. If your model is effective, KM plots for low and high-risk groups would have a significant difference (measured by log-rank test statistically).
• Calibration curve plots are often used to evaluated the performance of model that outputs probabilities or likelihoods ranged from [0.0, 1.0]. We can calculate the survival function, and then specify a time point of interest, e.g. 5-Year. At last, we compare the survival probabilities with the actual survival state at the specified time, which is just the same as we do evaluating a binary classification model. More details of obtaining survival function can refer to my tutorial, and the principles can be found in that post aforementioned.