The study actually answers a subtly different question than the prior literature. They measure the total amount spent on people who they predict to be in their last year of life, as opposed to looking back at people who are already known to have died (what they call ex-ante vs. ex-post analysis). Their conclusion is essentially that care in the last year of life is expensive, but it's hard to predict a priori which patients are trending in that direction.

They use a novel machine learning model to estimate time to death for patients in their dataset, but they don't seem to give a rigorous evaluation of its predictive accuracy. (They use AUC as the primary metric for a regression model?) So I'd take some of these results with a grain of salt.