Random forests require 10x the sample size as mostly additive regression models, so I’m failing to see a role here. I would rather use splines on all the continuous predictors, devoting the number of knots to each predictor in descending order of potential the predictor has. This strategy is described in detail in my RMS book and course notes.
Until you posted this and sent me looking, I hadn’t been aware of van der Ploeg, Austin & Steyerberg (2014) that discussed the data neediness of Random Forest.
Interesting paper, thanks for the comment. (I’m here to learn … it’s working.)