LLM4EHR Aligns Clinical Data for Improved ICU Predictions

Recent advancements in clinical machine learning for intensive care unit (ICU) outcome predictions have moved towards foundation models that leverage diverse clinical data modalities. However, existing methods often overlook the shared temporal structures between clinical events and time series observations within Electronic Health Records (EHRs), which can limit model robustness and adaptability. To address this, researchers propose LLM4EHR, a novel clinical foundation model trained on ICU EHR data. It combines domain-adapted large language models with a Transformer time series encoder, pre-training the model by explicitly aligning EHR events and time series data. A regularized contrastive objective helps learn robust time series representations conditioned on event embeddings from the LLM, leading to improved performance on various clinical tasks and demonstrating transferable embeddings for new patient cohorts.