Exploring optimal encoders for electronic health records

Abstract

Making the most use of abundant information in electronic health records (EHR) is rapidly becoming an important topic in medical domain.A recent work presented a promising framework that embeds entire features in a raw EHR regardless of its form and medical code standards.Despite its merit, the resulting embedded data is extremely large compared to the amount of EHR information it contains.In this paper, we explore for the optimal encoder not only reducing the large data into a manageable size, but also well preserving the core information of patients to perform clinical tasks. We found that CNN structured in a hierarchical manner outperforms the state-of-the-art model on widely accepted tasks in the field, even with fewer parameters and less training time. Moreover, it turns out that making use of the inherent hierarchy of EHR system can boost the performance for any kind of backbone models and clinical tasks performed.By conducting extensive experiments, we present concrete evidence for generalizing our research findings into real-world practice.We give a clear guideline on building the encoder based on summary of research findings captured while exploring numerous settings