Open viyolopez28 opened 2 years ago
Hello @viyolopez28,
Thanks for being interested in our paper.
Here's some of my understanding of your question:
1) Q: Should we remove the highly correlated radiomics features? A: Yes. If we have a lot of radiomics features but not too many data samples, it may be easy to have an overfitting problem. So we can remove the highly correlated radiomics features (for example, with an absolute correlation greater than 0.95) to decrease the number of features.
2) Q: Should we remove the highly correlated features before or after ComBat harmonization? A: -- Notice that, these two solutions will lead to different harmonization results. This is because, when using the parametric or nonparametric Empirical Bayes (EB) methods to solve the ComBat model, EB methods are designed to "borrow information across features" to make the parameter estimation more stable. --I think both solutions are Okay, but I would first try to remove the highly correlated features before ComBat harmonization.
Hope the answer helps you a little. Do not hesitate to leave a message if you have more questions.
Have a nice day! Best, Yingping
Hello, @Yingping-LI
I find so interesting the article
impact of pre-processing and Harmonization method on the removal of scanner effect in MRI radiomic features
. Assuming that I have radiomics features high correlated, should I remove them before harmonization?thank you.