Figure 6
From: EHR foundation models improve robustness in the presence of temporal distribution shift
The proportion of patients correctly and incorrectly re-classified by CLMBR-LR. Green bars indicate the proportion of patients that were incorrectly classified by count-LR but were correctly re-classified by CLMBR-LR. Red bars indicate the proportion of patients that were correctly classified by count-LR but were incorrectly re-classified by CLMBR-LR. Together, these represent the percentage of decisions that would have been affected if CLMBR-LR were used instead of count-LR. LR logistic regression; CLMBR clinical language model-based representations; LOS length of stay; ICU intensive care unit.
