I want to build a clinical prediction model and only can collect data from a small number of patients (let's say 500). Is it better to build a model by including only putative predictors in the model instead of everything that is available as often done in the machine learning community? I would think that using only 10 or 20 known/assumed predictors will produce a more robust model than including e.g. 1000s of gene data with no prior knowledge of its usefulness for this model. Is there any reference available?
I would assume that adding established and putative predictors in your main model and present the resulting model as your main model. You can then rerun the model with al predictors as an explorative analysis and identify further potential important predictors, which you then can further assess in a new study.