Unifying Heterogeneous Electronic Health Records Systems via Text-Based Code Embedding

Abstract

EHR systems lack a unified code system for representing medical concepts, which acts as a barrier for the deployment of deep learning models in large scale to multiple clinics and hospitals. To overcome this problem, we introduce Description-based Embedding, DescEmb, a codeagnostic representation learning framework for EHR. DescEmb takes advantage of the flexibility of neural language understanding models to embed clinical events using their textual descriptions rather than directly mapping each event to a dedicated embedding. DescEmb outperformed traditional code-based embedding in extensive experiments, especially in a zero-shot transfer task (one hospital to another), and was able to train a single unified model for heterogeneous EHR datasets.