I'll try not to beat around the bush. How would you approach, if you had imputed a lot of numerical features with k-Nearest Neighbors imputer and then wanted to get some MI scores on all of your features?
As you can't store mentioned numerical features back into your "old" dataframe (because you have to standardize the values first and the dataframe contains object features too) and you can't use your "old" dataframe to build the MI scores because of the various null values.
Trying to solve this for about 2 hours now, maybe I can't see the tree in the forest anymore :D
PS: I'd like the idea to impute with KNN (just learned about it) so I didn't want to bother with an alternative solution that gives my the same robustness.