There's two issues here:

The raw performance of the method itself is pretty bad. A lot of time is spent checking types:

This could be fixed by something like this: https://gist.github.com/smoogipoo/fda9c73371309a2a581c726a12bceeb9, which I'd say is something we can apply with little thought.

That change will already make it somewhat reasonable for one displayed combo counter. The rest is because it's running over every single HitEvent every judgement, once per counter. Perhaps this could be improved by storing incremental data locally or via ScoreProcessor, but there comes the question of handling rewind (maybe you don't and static calculation?).

Leaving this here for now as I'm reprioritising, for if anyone else wants to take the lead.