Hello, I'm conducting some experiments using TCGA-LUAD clinical and RNA-Seq count data. I'm building machine learning models for survival prediction (Random Survival Forests, Survival Support Vector Machines, etc.).
In several papers, I’ve noticed that differential expression analysis is often used as a first step to reduce dataset dimensionality. However, I’m not entirely sure how this step should be integrated into the modeling pipeline.
Specifically, should the differential expression analysis be incorporated within the cross-validation process?
My current idea is to select appropriate samples for the DE analysis (tumor vs. adjacent normal tissue), filter the genes based on the DE results, and then perform cross-validation experiments using this reduced dataset (excluding the samples used for the DE step, the tumor ones, since adjacent tissue samples are not used for model training).
Would this approach be correct? I’m concerned about potential data leakage if DE is done prior to cross-validation.
