Cross Validation not as accurate because I have NaNs in my training data

I am using N-1 cross-validation to train a vector to predict IQs based on connectivity values. Basically, my vector train_vec contains NaN vals. I wondering how I can fix this to get a better prediction. I have tried omitting but get the same result, tried filling the vectors with 'previous' or 'next'. If I remove this will affect the connectivities associated with a particular value.

IQ has a guassian distribution so is there any way to fill in the NaNs based on this?

New to matlab to still learning the coding...

Any help will be much appreciated