LLM4EHR: Aligning Clinical Time Series with Medical Event Sequences via Large Language Models

Short summary

LLM4EHR is a clinical foundation model that aligns ICU electronic health record time-series data with medical event sequences using domain-adapted LLMs and a transformer time-series encoder. The model is pre-trained with a regularized contrastive objective that learns robust EHR time-series representations conditioned on LLM-produced event embeddings. Ablation studies show improved downstream clinical task performance and transferable embeddings adaptable to new cohorts via k-shot learning.

• •Clinical foundation model aligning EHR time-series with medical event sequences via domain-adapted LLMs
• •Regularized contrastive pre-training learns robust clinical TS representations conditioned on event embeddings
• •Transferable embeddings support k-shot adaptation to new patient cohorts with competitive downstream performance

Generated with AI, which can make mistakes.