One Loss to Rule Them All: Marked Time-to-Event for Structured EHR Foundation Models
Researchers have introduced ORA, a novel pretraining objective for Electronic Health Record (EHR) foundation models. This method, called marked time-to-event, jointly models the timing of clinical events and their associated measurements. Unlike previous approaches that primarily focused on next-token prediction, ORA aims to capture the full complexity of EHR data, including continuous measurements. Experiments across various datasets and model architectures demonstrate that ORA yields more generalizable representations and improves performance on downstream tasks such as regression and time-to-event prediction. AI
IMPACT Enhances EHR foundation models by better capturing clinical event timing and associated measurements, potentially improving downstream healthcare predictions.