Electronic health records (EHR) contain data from disparate sources, spanning various biological and temporal scales. In this work, we introduce the Multiple Embedding Model for EHR (MEME), a deep learning framework for clinical decision support that operates over heterogeneous EHR. MEME first converts tabular EHR into "pseudo-notes", reducing the need for concept harmonization across EHR systems and allowing the use of any state-of-the-art, open source language foundation models. The model separately embeds EHR domains, then uses a self-attention mechanism to learn the contextual importance of these multiple embeddings. In a study of 400,019 emergency department visits, MEME successfully predicted emergency department disposition, discharge location, intensive care requirement, and mortality. It outperformed traditional machine learning models (Logistic Regression, Random Forest, XGBoost, MLP), EHR foundation models (EHR-shot, MC-BEC, MSEM), and GPT-4 prompting strategies. Due to text serialization, MEME also exhibited strong few-shot learning performance in an external, unstandardized EHR database.

Competing interests: J.C. receives research funding from Optum Labs. The other authors declare no competing interests.