Figure 1

An overview of the two approaches of constructing patient representations used in this study. The purple box in the construction of count-based representations represents the reference range comparison and binary feature construction procedures for a specific time-bin. The construction of CLMBR illustrates the self-supervised pretraining stage, hence the inclusion of the self-supervised learning objective. The adaptation of CLMBR to specific tasks (e.g., for predicting hospital mortality) does not include the self-supervised learning objective. In addition, during adaptation CLMBR weights were frozen, and a separate classification head is learned on the same patient representations for each clinical prediction task. CLMBR Clinical language model-based representations.