You are right. It would work well if the models included a utility function and, instead of measuring likelihood of survival, the measurement were expected utility.