Predicting Hospital Readmission after Cancer Surgery with Survival Analysis and Machine Learning

Random survival forests are very useful in the prediction the time to the first hospital readmission and to model the effects of predictors on survival data with no assumptions. Their use is consequently constantly increasing in medical research. However, the clinical interest may rely in the multiple readmissions patients undergo. As a result, we have adapted these forests to survival data with multiple event occurrences, referred as recurrent events, in the presence or absence of a fatal event. The aim of this work is to introduce the use of RecForestto help understanding the burden of disease after cancer surgery. Our approach is in five steps, to i) discern the relevance of multiple and fatal events, ii) develop trees and build the forest, iii) thoroughly evaluate performance, iv) provide the importance of variables, and v) enable predictions on new data. Data from hospital readmission in metastatic colorectal cancer patients were used to illustrate our purpose