Value-aware transformers for 1.5d data
Sparse sequential highly-multivariate data of the form characteristic of hospital in-patient investigation and treatment poses a considerable challenge for representation learning. Such data is neither faithfully reducible to 1d nor dense enough to constitute multivariate series. Conventional models compromise their data by requiring these forms at the point of input. Building on contemporary sequence-modelling architectures we design a value-aware transformer, prompting a reconceptualisation of our data as 1.5-dimensional: a token-value form both respecting its sequential nature and augmenting it with a quantifier. Experiments focused on sequential in-patient laboratory data up to 48hrs after hospital admission show that the value-aware transformer performs favourably versus competitive baselines on in-hospital mortality and length-of-stay prediction within the MIMIC-III dataset.
PDF Abstract
MIMIC-III
