Risk score stratification with k-means

Question

[julia.gilhodes]

Dear all,

I have created a risk score (point-based risk score) to predict time to brain metastases from covariates selected with a backward approach. My risk score ranges from 0 to 45. I need to classify my patients into 4 risk groups:low/intermediate low/ intermediate high/high. My question is: can I use a clustering approach like k-means, by including my clinical characteristics and my risk score, to obtain clusters with similar risk scores and clinical characteristics?

Thank you in advance for your help.

Answer 1

[Mensur Dlakic]

A general answer to your question is yes. In practice, it depends on how many features (clinical characteristics) and data points (patients) you have. It also matters what you plan to do with these clusters: are they just for internal bookkeeping or for future diagnostics? If you are planning to classify/diagnose new patients, I think it may be better to do a formal machine learning approach with (cross)validation, and test it on unseen data. That will also depend on the number of data points and informative features.

There is a great risk of overfitting if you simply cluster based on presently available data, especially if the dataset size is small.