Stimulus generalization comes to mind. IIRC Pavlows dogs at some point reacted not just to the food, but even earlier on him entering the room.

I doubt that they train dogs like "Steve's having a seizure. Good! Get that dog here."

I imagine that you do a preliminary training with faked seizures to get responses right and then hope (meaning: Train under real conditions) that the dog will react on a real seizure as well.