Categorizing Date in a registry data base of cancer

Question

I happen to work with a breast cancer database which is not available online. I want to follow these steps and report the results:

1. Remove data of people who are young( less than 18 yo) and who are old(more than 65 yo)
2. Replacing illegible data with NULL values
3. Imputing those NULL values with MICE
4. Feature selection using Rapid-Miner
5. Predicting death caused by breast cancer with SVM and Kaplan-Meier

Most of the features are categorical but some of them are numerical, I have solved some of the numerical ones except the dates which could help in the prediction.

For example: Surgery : Yes or No and if Yes when "Date is here"

I want all my features to be categorical because I think this way I would have better results. If you think otherwise please tell me and If you have any suggestions explain it to me, and also if you think i can skip all the dates and just continue without it let me know

I would be appreciated by any suggestion for all of those steps and if you have any method that I can be used for categorizing dates please HELP me out Thank you all

Answer 1

You will not be better off categorizing your continuous predictor variables. This is discussed in many places on this site, for example on this page. With continuous predictors you can evaluate their potential non-linear relationships with outcome while using up very few degrees of freedom in the analysis.

For display of survival curves you might want to choose example predictions based low versus high values of a continuous predictor, or directly plot the predicted hazard ratio and confidence intervals as a function of the value of the predictor. But your initial analysis will best be based on continuous values.

Specifically with respect to surgery dates, there are two major ways those might be used.

First, the date of surgery (and definitive pathologic diagnosis) is often used as time = 0 for survival analysis. If times of subsequent follow-up visits and events are expressed as calendar dates, then you need to know the date of surgery precisely to calculate event and censoring times for the survival analysis.

Second, there can be changes in therapy over time that affect survival. For example, there have been such improvements in technology for post-surgery radiation. You could then use the surgery date itself as a predictor, serving as a surrogate for improved therapies (and other time-varying factors) if you don't know exactly which technology was used in individual cases.

In either case, though, you would want to know the actual surgery date.