JMIR Publications

ABSTRACT

Background:

Transformer models of electronic health records (EHR) can accurately predict intensive care unit (ICU) outcomes, but most only use data collected during the ICU stay itself, and it is unclear whether a patient's pre-admission medical history can further improve that performance.

Objective:

This paper presents TransEHR2, an extension of Xu et al's TransEHR framework, and uses it to test whether historical medical records provide marginal predictive value over in-stay records for ICU outcome prediction using MIMIC-IV.

Methods:

TransEHR2 accommodates several additional data types beyond those supported by TransEHR: vector-valued numeric, ordinal, categorical, and text features, the latter enabling the use of discharge summaries. It also uses a reformulated transformer Hawkes process loss that models the joint likelihood of event types and timestamps rather than timestamps alone. We compared TransEHR2 to two baseline models. The first model only used records from the first 48 hours of the ICU stay to predict in-hospital mortality, length of stay, and diagnoses. The second model exclusively used records that predated the current ICU admission. The remaining experimental models were used to determine the marginal predictive value of historical records over in-stay records.

Results:

Pre-admission records had minimal predictive value compared to in-stay records. Historical records did not add any marginal predictive value over in-stay records alone in experimental models that used both, and models that incorporated historical text (discharge summaries, diagnosis descriptions) performed worst overall.

Conclusions:

For the evaluated features and prediction tasks, pre-admission history adds no value once records from the first 48 hours of the ICU stay are considered.