Feature importance (top features being one hot encoded features)

my feature importance report shows the top feature being a one hot encoded value. infact there are several in the top 10. when i split data used for training into my positive class (2%) and my negative i don’t see any major differences in percentage of these categories between the two.

how does xgboost split on categorical one hot encoded features? e.g. if it is ‘56’ say, at the tree how does it split?