Frank Harrell and many others say - dichotomization should be avoided for the power maximization, it can easily be checked using simulations: Wilcoxon test is more efficient than any other form of binning. That's the preamble.
But in clinics the response (e.g., cancer stage) is really often coded as interaction between clinical parameters .
1) How dichotomization can be avoided in this case disregarding the clinical classification that already exists? (i.e., we are interested in what makes tumor worse in general, no matter how clinicians determine it)
2) You have a dataset of infinite amount of cancer samples (heterogeneity is as small as possible, one cancer type is considered) and their overall survival / progression free survival / other outcomes. How would you divide the cancer into stages? The model is $Outcome \sim Features(size, numberOfMetastasis,etc)$. In other words, which unsupervised clustering method you would perform to determine groups that have different overall survival? Your goal is to do "staging", but using the power of AI and Machine Learning and Deep Learning =)
References to papers as answers are totally accepted, no need to describe the whole theory, I just don't know where to start the search process. Some sort of binning for decision making in the end seems unavoidable - many of the treatment options are binary (surgery / no surgery, radiotherapy with continuous dosage / no radiotherapy, etc).