The interpretation of the system input and output variables should remain the same. It is only the internal states that change. If you are directly using one of the states in the original model as if it were a system output, you should instead assign an output to this state, and then do your balanced reduction. The new system C and D matrices will compute the correct combination of the new systems original states, to produce an output that is equivalent to the state you were interested in in the original system.
So for example if you had a system with 6 states and were interested in the preserving the interpretation "meaning" of the third state in your system, you would add a row to your C and D matrices, as
Caug = [C;[0 0 1 0 0 0]]
Daug = [D;[0 0 0 0 0 0]]
And then use Caug, and Daug in your system which has its order reduced
