Interesting to see that they relied on random forests for feature selection. I'd wonder if a more "classical" approach would have yielded similar/comparable results, or if these findings where only correlatable due to the use of ML.