tl;dr binarizing mortality prediction makes models really sensitive to time horizons. models trained on acute care patients show more value than those from outpatient care, where performance is only slightly better when including ECGs. site-specific models are generally better than global ones.

Although a number of studies have shown decent performance from one-off models, we've done a more thorough look at how modeling decisions (e.g. classification versus survival prediction, choice of covariates, health setting) affect mode performance. the preprint is here: arxiv.org/abs/2406.170...