Similar to how we "know" how many people a low-level nuclear release killed: You know how much of a high dose of some carcinogen it takes for a certain cancer o

Similar to how we "know" how many people a low-level nuclear release killed: You know how much of a high dose of some carcinogen it takes for a certain cancer or death risk. You scale this down linearly to low doses. Then you estimate population doses from the release and a few multiplications later you arrive at a "death toll". You will note the amount of estimates necessary, and the pessimistic assumptions like linear-no-threshold for carcinogen exposure. Most probably, death-tolls estimated this way are upper bounds. Only for smog events, there is a possible calibration of such estimates through excess mortality, and even that is in some doubt, because smog hits the old-infirm-dying-anyways harder.